Real-Time Adversarial Reasoning Under Uncertainty
Real-Time Adversarial Reasoning Under Uncertainty
Proposal to DARPA Program on
Real-Time Adversarial Intelligence and Decision Making (RAID)
Submitted to Dr. Alexander Kott, IXO
Wednesday 21 April 2004
Volume I: Technical
BAA Number:
04-16
Assigned DARPA Control Number:
P-0416-100363
BAA Category:
Adversarial Reasoning
Proposer Point of Contact
Dr. William H. Hsu
Department of Computing and Information Sciences
Kansas State University
234 Nichols Hall
Manhattan, KS 66506-2302
Voice: (785) 532-6350 ext. 29
Fax: (785) 532-7353
E-mail: bhsu@cis.ksu.edu
Table of Contents
A. Table of Contents 2
B. Proposal Roadmap 4
C. Problem Statement 5
C.1 Challenges of Predictive Analysis 5
C.1.1 Urban Complexities 5
C.1.2 Decision Analysis in Complex Urban Environments 6
C.1.3 Automated Course of Action (COA) Development and Analysis 6
C.1.4 Determining and Displaying Results with Maximum User Benefit 7
C.2 Limitations of Current Approaches 7
C.3 Opportunities for Improvement 8
C.3.1 Benefits to the Military: Impact of Computational Analysis 9
C.3.2 Scientific Benefit 9
D. Program Concept 10
D.1 Proposed Enabling Capabilities 10
D.1.1 Course of Action (COA) Generator for Urban Operations 10
D.1.2 Dynamic COA Analysis 11
D.1.3 Bayesian Learning and Reasoning for COA Assessment/Selection 13
D.1.4 Multi-Agent Reasoning 14
D.2 Proposed Capability Development 15
D.2.1 Adversarial Reasoning Module (ARM) Overview 15
D.2.2 Prediction and Detection of Changes in Context 16
D.2.3 Real-Time Decision Making 16
D.2.4 Decision Support and Expert Overrides 17
D.2.5 Integration with Deception Reasoning 19
D.2.6 Results Determination and Display 19
D.3 Proposed Performance Metrics 20
D.3.1 Operational Metrics 21
D.3.2 Technical Metrics: Decision Making and Predictive Validation 21
E. Technical Approach 22
E.1 Survey of the Current State of the Art 22
E.1.1 Related Work 22
E.1.2 Relational Data Mining (NSF-Funded Effort, 2002-present) 26
E.1.3 Real-Time Decision Support (ONR-Funded Effort, 2000-2002) 27
E.1.4 Simulation-Based Optimization (ONR-Funded Effort, 1999-2002) 29
E.1.5 Performance-Based Measures of Skill in Microworld Simulations (NSF, FAA & ONR-Funded Effort, 1996-1998) 32
E.1.6 Immersive Training Systems for Crisis Management (ONR-Funded Effort, 1995-1998) 32
E.2 Approach for Technology Development 32
E.2.1 Data Modeling: Ontology and Representation 33
E.2.2 Algorithms for Automated Learning and Reasoning 33
E.3 Key Ideas for Future Development 33
E.4 Self-Evaluation Methodology 34
F. Management Approach 35
F.1 Statement of Work 35
F.2 Program Schedule 36
F.3 Deliverables 37
F.4 Cost Summary 38
F.5 Personnel 43
F.6 Related Experience 47
F.7 Facilities and Teaming 48
F.8 Security Plan 50
F.9 Statement of Rights Claimed for Software Deliverables 50
G. Self-Assessment According to Evaluation Factors 51
G.1 Technical Depth and Feasibility 51
G.2 Consistency with RAID Program Concepts 51
G.3 Cost realism 51
G.4 Personnel and Corporate Capabilities and Experience 51
Proposal Roadmap
This proposal describes a plan to develop the adversarial reasoning module (ARM) of a real-time adversarial intelligence and decision making (RAID) system that will interoperate with an urban combat simulator for small cohorts. Beginning with a problem definition (Section C) in this domain, it presents a research and development program that will produce the required capabilities (Section D). In Section E, related methodologies are surveyed. Then a technical research program is described and novel contributions are reviewed. Finally, a management plan in Section F describes the project deliverables, intermediate milestones, and development timeline.
Goals of ARM: Urban mission planning against an opposing force, with hidden information, poses a challenge in autonomous agent reasoning and learning. General problems in this domain require consideration of multiple sources of uncertainty:
• enemy agent: location, identity, condition, capability, information state, plans, and execution stage – our prediction and reasoning goals
• terrain: accounting for visibility, fire, cover, dispersal of friendly entities
• resource availability (e.g., ammunition) for red and blue teams
• urban doctrine: reasoning with background knowledge
• logistics: time/resource bounds, especially accounting for resupply, reinforcements
Our goal is to develop probabilistic representations to support automated reasoning and learning from data produced by interactive combat simulations.
Overview of R&D Plan: Our plan consists of three overlapping development phases corresponding to aspects of our system design:
• Test bed – data models, ontology, event communication language, reasoning and learning framework; desktop PC platform, portable (Java-based) infrastructure
• Online reasoning and learning algorithms – representations for uncertain aspects listed above; algorithms for detecting and adapting to change in hidden context
• Interfacing and experimentation – cooperation with deception reasoning module (DRM) development team and system integrator to build overall RAID system; cooperation with experimental evaluator to assess performance gain from ARM
The proposing team consists of three institutions, a prime contractor and two subcontractors, with complementary capabilities that support ARM development:
• Kansas State University: probabilistic reasoning, machine learning, (3-D) spatial modeling, databases, real-time decision support (Computer Science); interactive microworld simulations, judgement / decision making, culture / doctrine (Psychology)
• University of Mississippi: simulation-based optimization, decision-making algorithms, Bayesian analysis; adversarial reasoning
• American Systems Corporation (Virginia): data integration and modeling; user interface development; subject matter expertise in urban doctrine, tactics, logistics
Problem Statement
1 Challenges of Predictive Analysis
The Real-time Automated Intelligence and Decision-making (RAID) program poses two kinds of challenges that affect decision analysis. First, urban operations by their very nature offer a range of complexities—complicated physical terrain; extensive physical and service infrastructure; and a multiplicity of interactions among a large, dense population —that must be taken into account when making decisions. Second, and key to the development of an effective Adversarial Reasoning Module, there are important technical challenges associated with automating the development and assessment of Blue, Red, and Green (or White[1]) courses of action in this complex environment.
1 Urban Complexities
Urban doctrine describes an array of factors that commanders, their staffs, and troops must address. These factors are sometimes depicted as an urban triad:
▪ Complex, manmade physical terrain that encompasses city patterns (geographic features, neighborhoods, street grids, etc.); buildings of varying types and sizes; a blend of horizontal, vertical, interior, and exterior areas; and areas consisting of airspace, surface, supersurface, and subsurface elements.
▪ Population of significant size and density that may include combatants, other potential threats and actors, as well as non-combatants, with each having various affiliations, demographics and interactions
▪ Physical and service infrastructure that at a bare minimum includes a transportation network (streets, tunnels, rail lines, etc.); communications grid; utilities; distribution of food, water, and other goods and services; police, fire, and other emergency and social services.
These and other urban characteristics have notable effects on capabilities and consequently, on decision-making, which must take into account own force capabilities, adversary capabilities, and the current situation as perceived at a given point in time. The reasoning module will have to address the diverse array of characteristics that define urban operations and affect capabilities, such as:
▪ Ability to sense may be affected, with C4ISR degraded and forces more dependent on line-of-sight capabilities. Three-dimensional (3-D) visibility constraints affect targeting and air-based intelligence gathering.
▪ Communications interference, background noise, and connectivity issues may introduce disruptions and delays in electronic and/or oral communications.
▪ Command and Control (C2) in this environment may lead to longer decision cycles (due to the time required to acquire, process, and disseminate information) and result in more decentralized operations
▪ Cohesion/Mass may be reduced by the dense three-dimensional nature of the environment, terrain features, barriers, etc. Large numbers of troops may still equate to low density.
▪ Movement/Maneuver will take place in a three-dimensional environment – inside and outside as well as above, on, or below the ground – with channelization, terrain features, urbanization, obstacles, etc. influencing decisions and actions. Damage from tactical explosives and missiles is also affected by terrain.
▪ Employment of Fires may be constrained not only by the available organic options but also by engagements characteristics, operational constraints, and policy considerations.
▪ Time/Speed factors will reflect changes in rates and the effect of distances, areas, and volumes to be covered.
We see the greatest technical challenges to predictive analysis taking three forms: 1) Decision analysis in a complex, urban environment; 2) Automated course of action (COA) development and analysis; and 3) Determining and displaying results with maximum user benefit.
2 Decision Analysis in Complex Urban Environments
Warfare in general and urban operations in particular are complex, non-linear phenomena. Consequently, outcomes are sensitive to initial conditions, and results will be probabilistic, not deterministic. Decision analysis will have to account for the assessment of options based on partial information, different sets of partial information being available to different parties, each party’s reasoning based on available information and perceptions. Moreover, the ARM will need to maintain sets of perceived data (as well as ground truth from which outcomes can be calculated) in order to provide feedback that alters perceptions over time and the underlying drivers of decision-making. Incorporating feedback and the perceived results of prior actions will be challenging but key to offering an adaptive reasoning capability.
3 Automated Course of Action (COA) Development and Analysis
Operational users have great familiarity with developing and assessing courses of action (COAs). The process for analyzing a situation and mission to determine specified, implied, and essential tasks; constraints and restraints; etc. is well understood. Operational users begin this process, however, with an understanding of the situation. The ARM must automate this process in a way that’s applicable for any situation.
Automating the development and analysis of COAs will start from a baseline—derived from doctrine; tactics, techniques, and procedures (TTPs); case studies; etc.—and build on this by incorporating outcomes from earlier actions (and the perceptions of the actions and outcomes from the perspective of Blue, Red, and Green/White). Another key aspect necessary for automated assessments involves scoring. This will enable alternative COAs to be compared and will play a key role in addressing how Blue and Red choose to act.
The power of automating COA development and analysis will be manifest in breadth and depth that no individual (or staff) could manage.
4 Determining and Displaying Results with Maximum User Benefit
As noted, given the nature of the decisions/actions being analyzed, outcomes will be probabilistic rather than deterministic. The ARM will make use of Monte Carlo simulation to derive meaningful statistics, presented in a way that makes it easy for a user to understand and act on. The determination and display of results will also involve the integration of Payoff and Risk information and will highlight potential insights relevant to the user (such as changes in key patterns and underlying rules).
2 Limitations of Current Approaches
Current real-time decision making representations and algorithms do not adequately account for certain aspects of urban combat, real or simulated.
• Resource-constrained reasoning and learning: Hard and soft real-time deadlines corresponding to real-world utility loss in a combat scenario are difficult to meet using current state-of-the-art algorithms for inference and mapping. Recent work by Dechter et al., Koenig, Zilberstein, etc. on adapting incremental reasoning algorithms to meet various real-time requirements was surveyed by the principal investigator. These algorithms perform search in uniformly-partitioned space and inference using representations such as graphical models of probability (e.g., Bayesian networks and dynamic Bayesian networks aka DBNs). Time-space tradeoffs for Bayes net inference may not guarantee graceful degradation of prediction quality, one of the requirements of the proposed ARM.
• Multiple sources of uncertainty: On the other hand, representations and algorithms that support graceful degradation under computational time and space limitations may not be sophisticated enough to cope with the many sources of uncertainty listed in section C.1. The “urban triad” of artificial terrain, population, and infrastructure breaks this uncertainty into causal aspects. Similarly, complicating factors such as sensor obstruction and interference, policy, intelligence and logistics add to this uncertainty. Current representations such as probabilistic relational models (PRMs) provide a powerful means of abstracting and encapsulating this uncertainty, but they do not scale up well to domains such as urban operations, especially when it is necessary to represent time in the simulation and in the RAID application.
• Modeling time: Currently, time series models such as dynamic Bayes nets make many limiting assumptions (e.g., fixed time lag between event frames) that may be too limiting for the ARM application domain. More flexibility is needed in the representations used to capture time (e.g., sparse representations for long time lags).
• Continuous variables: As the proposal solicitation (DARPA BAA 04-16) notes, the domain is continuous rather than discrete – this may include not only continuous spatial distributions (e.g., probability distribution over the location of an enemy entity) and continuous time but inferred variables such as assessment of a threat level.
• Unified data representation for agent communication and adversarial reasoning: “Know the enemy and know yourself”, in this domain, requires a common description language for agents. To make predictions about the red team, red and blue must be represented within a common data model, ontology, event communication language. Though there are OWL (formerly DAML+OIL) ontologies for miilitary domains, a comprehensive one for urban operations against small cohorts does not yet exist. Such an ontology is clearly needed for the ARM and DRM, but developing one would potentially benefit other tactical planning domains.
• Situational awareness: Uncertainty about adversarial information state and plans needs to be modeled not only for deception reasoning but in order to generate predictions about actions such as deployment of noncombatant (green team) sympathizers as intelligence-gathering agents. The ARM must be able to answer: “What does the agent know?” for red, green, and blue team members in some contexts such as predicting likelihood of an ambush, detection through reconnaissance, and countermeasures.
• Spatial modeling: 3-D modeling will be required to account for entity movement, line-of-sight, and the success probability of attacks such as sniping.
3 Opportunities for Improvement
Aims and Scope: The focus of this collaboration among real-time intelligent systems, optimization, and judgment and decision-making researchers is development of an ARM for DARPA’S RAID requirement in the domain of urban combat. Adversarial reasoning requires situated learning and reasoning algorithms that can detect hidden changes in context and respond accordingly. The overall problem is as follows:
The target functionality is to predict actions committed by an enemy force (red team) consisting of 3-7 enemy agents, with potential noncombatant sympathizers who may relay intelligence information to the enemy.
The performance element must recommend one or more courses of action to the user, the CO of a dismounted company-sized friendly force (blue team) with local armor and air support.
All members of the proposing team will apply their experience with real-time inference and learning, data modeling and integration, and spatial reasoning to achieve this functionality. Some aspects of the abovementioned limiting factors and assumptions will thus be overcome.
Through automation, it will be possible to generate significantly more COAs than could otherwise be developed. These COAs, individually and in combination, will allow for many more factors of the situation (as it actually exists, as Blue perceives it, and as Red perceives it) to be assessed and reported. Automating COA development will also make it possible to account for alternatives based on changes in objectives, different policy and operational constraints, etc. This increase in breadth would yield higher Blue payoffs (because optimization would take place over a larger set) and would reduce the possibility of Red surprising Blue.
1 Benefits to the Military: Impact of Computational Analysis
In addition to increasing breadth, automating COA development and analysis would also promote assessments of much greater depth. Automation will allow for the systematic consideration of many more factors than would be possible during human assessment and wargaming of alternative COAs. Moreover, the automated ARM will support the analysis of multiple moves, allowing for more branches and sequels to be identified and evaluated. Finally the analysis will yield insights into underlying patterns and rules that guide decision-making.
Automation could also make a meaningful difference in the display of results. This may take the form of summary scorecards that address analysis across numerous COAs and branches/sequels. Well-designed scorecards would facilitate effective information processing and decision-making. Also, high Payoff pathways could be highlighted—at the level of initial COAs, anticipated action-reaction-counteraction sequences, and subsequent branches and sequels—so users could investigate them in greater detail. Finally, to support real-time cost-benefit decisions, the ARM we develop would present integrated Payoff and Risk information.
2 Scientific Benefit
Dynamic graphical models of probability can provide a representation that addresses the urban complexities listed in Section C.1.1, while combining current and new algorithms for real-time reasoning, machine learning, decision support, time series prediction. We hypothesize that this is necessary for development of an ARM capable of driving the recommendation function of the RAID system, and propose to demonstrate this piecemeal by introducing functionality incrementally.
Where the holistic nature of some of the functionality does not permit ablation studies (comparison between the ARM with full functionality and a variant with partial functionality such as artificially degraded accuracy), a baseline system consisting of rule-based predictors shall be used.
We also hypothesize that hidden changes in context can be captured for active learning techniques such as active reinforcement learning, by extending temporal graphical models of probability such as dynamic Bayes nets. One further contribution of this research shall be the novel combination of active learning algorithms – for inducing latent variables – with existing techniques for detecting hidden changes in context.
Program Concept
1 Proposed Enabling Capabilities
The Adversarial Reasoning Module will identify potential actions – encompassing the full range of available methods, means, and targets – and will assess and highlight the best options available to Blue.
Automating the development and analysis of and courses of action (COAs) will involve work that addresses the actual situation, Blue perceptions of the situation as well as the Blue assessment of Red’s perceptions, and Red perceptions of the situation as well as the Red assessment of Blue’s perceptions. Outcome calculations will be based on ground truth, but decisions will be made based on Blue and Red perceptions (with these perceptions being modified over time by the outcome feedback each is able to obtain).
This section describes the development of automated ARM capabilities as follows:
▪ Proposed Enabling Capabilities – This subsection focuses on four major areas: an Operationally-Based COA Generator, Dynamic COA Analysis, a Bayesian Approach to COA Assessment/Selection, and Results Determination and Display.
▪ Proposed Capability Development – This describes the proposed sequence of development for the major proposed capabilities.
▪ Proposed Performance Metrics – The section concludes with an identification of operational and technical metrics.
As discussed, in order to develop the Adversarial Reasoning Module, it will be necessary to automate the process for identifying available options and COAs, analyzing these COAs, scoring alternatives in a way that accounts for all players’ perceptions and addresses multiple-move pathways (looking at action-reaction-counteraction sequences and branches and sequels), and determining and displaying results in a way that meets the needs of the user.
1 Course of Action (COA) Generator for Urban Operations
The automated identification and development of courses of action will have a baseline that draws from past experience and a dynamic component that introduces feedback from previous RAID actions, outcomes, and their impact on current options and decision-making.
Baseline development will make use of reference information (both U.S.- and OPFOR-related) on command and control (C2) and decision-making. Subject matter experts possessing real-world operations experience across the spectrum of conflict and having expertise in urban operations and small-unit tactics will examine doctrine, TTPs, case studies, lessons learned, and other relevant information. This work will lead to a baseline set of rules and a detailed description of the decision-making process.
The most significant element of automating COA development will involve mission analysis across numerous specific situations to ensure that the process for identifying and characterizing distinct courses accords with sound military judgment and can be applied to any situation. Operational users have great familiarity with developing and assessing courses of action (COAs). The process for analyzing a situation and mission to determine specified, implied, and essential tasks; constraints and restraints; etc. is well understood. An individual or staff begins this process, however, with an estimate of the situation. Our subject matter experts will develop an appropriate way to automate this process so that it’s applicable for any situation and can be translated into algorithms. This will lead to software that generates a broad range of realistic courses of action for any situation.
Automating the generation and analysis of COAs will reflect both baseline information—derived from doctrine; tactics, techniques, and procedures (TTPs); case studies; etc.—and outcomes from earlier RAID actions (and the perceptions of the actions and outcomes from the perspective of Blue, Red, and Green/White). This feedback will make COA generation a dynamic process that allows for the incorporation of results and lessons (as well as the evolution of patterns, rules, and perceptions) and that makes the ARM adaptive.
The ARM’s automated COA Generator capability will also support the identification of branches and sequels, as the outcome of one action-reaction-counteraction sequence will lead to the generation and analysis of follow-on COAs. This multi-move analysis will generate entire pathways that can be analyzed, scored, and optimized both with respect to maximizing Payoffs and minimizing Risks.
2 Dynamic COA Analysis
The generation of numerous COAs has value in and of itself, because this can reduce the likelihood of Blue being surprised by Red. Additional COAs can also yield higher Payoffs or lower Risks because optimization takes place over a larger set. Dynamic COA analysis, however, can yield additional benefits.
Dynamic COA Analysis in the ARM – whether addressing current actions or looking multiple moves into the future along various event pathways – will have to begin by accounting for differences between actual and perceived views for all player combinations as illustrated in the figure below[2].
[pic]
Figure 1. Accounting for the actual situation and different perceptions.
The actual situation must be tracked because the outcomes of Red-Blue engagements need to be calculated based on ground truth. Player perceptions must be tracked because this affects choices from available COAs.
The following figure offers an illustration of how COAs will have to be analyzed. At a given point in time, both Blue and Red will have a certain number of options available (although their respective number of options will usually differ). In this case Red chooses an action attempting to achieve a favorable outcome and Blue reaction. The COA choice is based on Red’s understanding of the situation and Red’s view of Blue. Blue may react in a way other than intended (or desired) by Red. This may be the case simply because Blue’s perception of the situation differs from Red’s. It may also be the case, however, because the ARM provided insights on possible Red actions. This may allow Blue to react (or preempt) in a way that benefits Blue (although the outcome will be based on the actual situation, meaning it may differ from Blue’s expectation because, in Phases II and III, neither Blue nor Red possess “ground truth” regarding the actual situation).
Figure xx. Accounting for the actual situation and different perceptions.
[pic]
The ARM will also address the further complication of possible Green actions (whether independent or on behalf of another player) and the use of White or impact on White. This will affect the perceptions of all players and the outcomes of actions (and the perceptions of outcomes).
3 Bayesian Learning and Reasoning for COA Assessment/Selection
Scoring is critical to conducting automated analysis and selection of courses of action and calculating meaningful outcomes. In developing the ARM, our team will include subject matter experts in Bayesian analysis and in urban warfare and small-unit actions. Together, these experts will develop a scoring system that provides operationally relevant results. At any given point in time, the set of Blue-Red pairings can be assessed using a Bayesian approach.
Our specific technical objective is to develop a metareasoning system that learns to detect hidden changes in context and uses this knowledge to as a guide to replanning or decision-making in a situated agent that interleaves planning and execution. As an example, let us consider a terrain mapping example: a agent might change modes in response to context changes. This could mean changing modes in response to variable terrain; taking shelter from red team fire or in response to environmental conditions such as heavy fog; or selecting and adjusting specific behaviors such as attempting to collect intelligence, depending on the agent’s self-estimated condition and inferred hazards given this context. In a multi-agent problem, such decisions include revising individual beliefs, which drives team members to communicate about joint intentions and responsibilities in order, which in turn updates shared goals and plans.
In our recent work, we have explored the use of probabilistic representation, learning, and inference in autonomous robotics. Our focus has been on adapting our previously successful techniques such as layered learning to models that can represent hidden context and facilitate automated discovery of contextual factors. Recently, our laboratory has begun to develop a suite of software tools for simulation and visualization of autonomous and mixed-initiative exploration tasks. The learning components of this system shall build upon an existing open-source reuse library for Bayesian networks called Bayesian Networks in Java (BNJ) that we have been developing and distributing since 2000 and has registered users at over 50 other research institutions.
4 Multi-Agent Reasoning
In the past few years, the broad usefulness of multi-agent systems in finding solutions to optimization problems has been recognized.
In our emerging multi-agent world, we should expect the broad use of multi-agent system addressing discrete optimization problems, according to Baker. Baker also comments that contrary to popular belief, most discrete optimization problems which can be globally optimally solved, can be globally optimally solved by a meaningful multi-agent system often in less time than the corresponding centralized algorithm. “Meaningful” in this context specifies a heterarchical architecture (i.e. a pure peer-to-peer system) where
• the computation across agents is substantially balanced,
• agents represent physical entities in the system, and
• agents can be dynamically a reproduction, reaction, and communication. dded or subtracted from the system with only polynomially sized changes in the algorithms which are being executed.
Liu and Yin stated that multi-agent approaches have shown to have a great potential in solving problems that are otherwise difficult to solve. Based on the idea of branching the authors designed a reactive behavior-based, distributed multi-agent approach to solving an integer-programming problem. The general design of the multi-agent system contains three key elements: a goal of the system, agent environments, and a behavioral repository for the distributed agents.
With the development of computers, researchers and practitioners have been able to deal successfully with larger and more complex problems. New paradigms of problem solving have been applied involving Artificial Intelligence, of which one of the most often used is the multi-agent system.
A lot of agent definitions can be found in the literature but there is no one which has been fully accepted by the scientific community. Moreover, the concept multi-agent has a higher frequency in the specialized literature. Some characteristics of agents are suggested as follows.
Wooldridge and Jennings define an agent by its flexibility, which implies that an agent is:
• Reactive: an agent must answer to its environment.
• Proactive: an agent has to be able to try to fulfill his own plans or objectives.
• Social: an agent has to be able to communicate with other agents by means of some kind of language.
Similarly, Drogoul, Vanbergue, and Meurisse describe an agent as an autonomous, proactive, and interacting entity.
These features can be observed in the example of multi-agent system presented by Davidsson, which is applied to the monitoring and control of intelligent buildings. Each agent corresponds to a particular entity of the building, e.g., office, meeting room, corridor, person, or hardware device. The behavior of each agent is determined by a number of rules that express the desired control policies of the building conditions. The occurrence of certain events inside the building (e.g., a person moving from one room to another) generate messages to some of the agents that trigger some appropriate rules(s) (Reactive). The agents execute the rule(s), with the purpose to adjust the environmental conditions to some preferred set of values (Proactive). The rule causes a sequence of actions to be executed, which involves communication between the agents of the system (Social).
Many agent architectures, at least theoretically, use rational choice for decision making, whereby all possible choices are “scored” and the highest ranked is chosen. Evidence suggests that people rarely use this type of decision making. Norling, Sonenberg and Ronnquist developed BDI (Belief-Desire-Intention) agents. These agents have beliefs, goals, and plans. Also, They created a limited form of memory about previous success of a plan, and by manipulating the rank calculation, the agent can exhibit some human-like qualities for plan selection when the context provided by the system designer is insufficient for this choice. However, the information captured is rather limited – it fails to capture why the plan has failed.
2 Proposed Capability Development
1 Adversarial Reasoning Module (ARM) Overview
The Adversarial Reasoning Module will identify potential actions – encompassing the full range of available methods, means, and targets – and will assess and highlight the best options available to Blue.
Automating the development and analysis of COAs will involve work that addresses the actual situation, Blue perceptions of the situation as well as the Blue assessment of Red’s perceptions, and Red perceptions of the situation as well as the Red assessment of Blue’s perceptions. Outcome calculations will be based on ground truth, but decisions will be made based on Blue and Red perceptions (with these perceptions being modified over time by the outcome feedback each is able to obtain).
The primary goal of this 36-month project is to develop and apply uncertain reasoning, multi-agent systems, and simulation-based optimization to real-time adversarial reasoning. The adversarial reasoning to be developed must respond to hidden changes in the context of information gathering by learning, autonomous agents.
2 Prediction and Detection of Changes in Context
Hidden change in context presents a fundamental and pervasive challenge to fault-tolerant computing and intelligent systems, particularly those that employ machine learning to explore a new environment. We focus on online learning because many real-world exploration scenarios such as mapping require it, and on RL-based approaches because they capture the tradeoff between exploitation and exploration of new environments. Furthermore, we propose to investigate dynamic graphical models because detecting hidden changes in context from spatiotemporal data often involves time series understanding and prediction – a problem for which DBNs admit effective representation, as well as reasoning and learning, techniques.
3 Real-Time Decision Making
The optimal sequence of interrelated tasks in order to finish a project by a certain time and within a cost constraint is a routine task when resources can be acquired at will. But when resources are limited, this presents a formidable scheduling problem. The list of tasks to be accomplished in a terrorist threat can be thought of as a project. This list may be small in comparison to a major building project but the scheduling must be done immediately and the task list and location of assets varies with each threat. Also, survival may depend on scheduling tasks in the correct order.
Our plan is to investigate constraint programming as a rapid solution technique for this important problem. Constraint Programming is the study of computational systems based on constraints. The earliest ideas leading to CP may be found in the Artificial Intelligence area of the Constraint Satisfaction Problem (CSP), dating back to the 1960’s. It is important to note that there is a difference between the word “programming” in mathematical programming and constraint programming. In CP, the word “programming” refers its roots in the field of a computer programming language. Other programming paradigms are procedure programming, objective oriented programming, functional programming and logic programming. In mathematical programming, the word “programming” roots in Dantzig 1963, which is associated with a specific mathematical problem.
The first modern constraint programming languages were extensions of logic programming languages, called constraint logic programming language (CLP). Three different groups independently developed CLP during late 1980’s and early 1990’s. In Melbourne, Joxan Jaffar et al, developed CLP[R]; In Marseilles, Alain Colmerauer et al developed Prolog-III and in Munich, at European Computer Industry Research Center, CHIP is developed. Another step toward CP came from the area of concurrent logic programming. One popular example is the Oz system developed by Smolka et al. Now there is the C++ and Java based object oriented constraint solver developed by ILOG, which integrates the constraint programming technique with the OPL programming language.
A CP approach to solve the RCPSP has several advantages. First, constraint satisfaction technique is good at handling binary constraints, which populate the RCPSP problem. In fact, binary constraints are ubiquitous in scheduling problems and since constraint programming is efficient at dealing with binary constraints we expect it to substantially reduce the solving time in this class of problems. Preliminary investigations show reductions by factors of 10 in small problems.
In particular we plan to make use of a product called OPL in this investigation. OPL is a computer application for constraint programming. OPL is the newest of these tools and promises to be effective in precisely this problem. In particular we plan to:
1. Construct a scheduling problem with characteristics similar to the problem of scheduling tasks during a terrorist threat.
2. Implement the problem in OPL as a constraint programming problem and in CPLEX as a mixed integer programming problem.
3. Compare the results of solving the problem under a variety of circumstances.
Our goal is to achieve the fastest method of scheduling tasks in real time to respond to a terrorist threat.
4 Decision Support and Expert Overrides
There are thousands of intelligent systems in use today that have been accepted and used on a regular basis on experts (T. Levitt). That some experts may override or turn off some systems on some occasions does not in any way suggest that humans can do without these systems. In an increasingly complex world, there is no choice but to rely on such systems. For instance, our most modern aircraft cannot fly without the operation of highly sophisticated systems (R. Schachter).
An ancillary objective of this work is to explore why experts users sometimes find such systems at odds with their human processing capacities and needs. If we can identify reasons for the mismatch between system design and human processes, then we can make concrete suggestions for better design and implementation of these systems.
We will now turn to a discussion of some of the possible psychological reasons why intelligent systems may be overridden or turned off by their human operators. These reasons are listed alphabetically rather than in order of importance:
#1: Attention Overload. It is a well-known psychological phenomenon that information overload can impair performance, even for experts. As a result, experts are often defined by their ability to simplify complex problems (J. Shanteau). They have no tolerance for irrelevant or partially relevant information sources. For instance, we have observed experts dismissing novices as a needless distraction in difficult situations. Even the best-designed intelligent systems require some attention from users, eg, by alarms or flashing lights. Since these can be a distraction, they are often ignored (or turned off).
#2: Conflict of Mental Models. The mental model or problem representation is critical to decisions made by both humans and computers (R. Thomas). These are usually in congruent in normal situations. In less straightforward settings, however, there can be a conflict between the mental model of the user and the problem representation designed into the automated system. This mode awareness mismatch can lead to seriously compromised decision making that the user resolves by turning off the computer system (K. Smith).
#3: Cost/Benefit Analysis. In many situations, the expert must decide whether it is worth the time or trouble to use (or even to learn about) an intelligent system. Often a quick but approximate answer is preferable to a slower but optimal answer (D. Scheid). In effect, the user performs a cost/benefit analysis and decides that the immediate cost in time and effort is too high given the potential benefit of optimality.
#4: Delayed Feedback. Due to unavoidable delays in feedback, intelligent systems are often “behind the curve.” For instance, it can take several minutes to download weather information to pilots. This can produce a conflict between the short-term tactical needs of the user and the long-term strategic goals designed into the system (J. Uhlarik). Even without such delays, it is difficult to design systems that look ahead by extrapolating into the future (J. Crow). That is, systems look back and users look forward.
#5: Design to the Mean. IS are designed by computer scientists, not domain experts (K. Vicente). Scientists solve problems they know how to solve and will attempt to maximize the hit rate of system performance by designing for typical cases. As a result, systems are built to deal with the average. But these are not necessarily the tough, but rare, cases (T. Levitt). And it is those cases that can lead users to disregard the system.
#6: Exertion of Control. Experts define themselves by their ability to take charge, ie, to exert control especially in extreme cases. When the situation is routine, however, experts are not challenged and thus are often uninterested; they do not care if a system takes over. For tough situations, experts want control – no guts, no glory (D. Gustafson). This can lead experts to override or turn off the system.
#7: Legal/Ethical Responsibility. Most intelligent systems (1) contain a well-specified set of operational rules, and (2) provide the ability to reconstruct events leading up to the decision, ie, they are transparent (Pingenot). This allows for reconstruction of the decision process. But it also opens the door for legal and ethical challenges, especially if the decision led to an undesirable outcome. Fear of lawsuits, for example, has led to the purposeful dumbing down of certain weather guidance systems. In turn, this can make users reluctant to open themselves up to newer systems.
#8: Miscalibration. IS use optimal combination rules, such as Bayes, that often yield more extreme results than human decisions (W. Edwards described this as conservatism). When humans see the output from such systems, they often recalibrate their inputs for routine cases. However, experts rarely know how to recalibrate for more unusual cases. As a result they may be reluctant to use such systems in extreme settings.
#9: Mistrust. Inevitably all systems designed by humans will fail, either due to equipment malfunction, design conflict, human errors, or other unforeseen problems. These crashes seem to occur at the most critical times – a perverse example of Murphy’s Law (if it can go wrong, it will.). If system failure is viewed as inevitable, it may make sense to turn off IS systems before they can cause severe problems.
#10: Perceived Risk. Incorporating intelligent systems into the decision loop inevitably leads to some increase in the complexity of the process. However, complexity is one of the factors that can increase perceived risk (R. John). It is well known that most humans are risk adverse and will take steps to reduce risk. This can be accomplished by overriding the output of an intelligent system.
#11: Personality: As demonstrated in the movie Top Gun, expert performers often display great confidence, even arrogance, in their own abilities. It is part of their mystique (J. Shanteau). The culture and personality of experts is to control inanimate objects, ie, computers, not to credit them for any successes. IS not only challenges this mystique, they can make the experts feel like novices – a most undesirable situation for any expert.
Research Questions: First, which of these factors, either individually or in combination, is responsible for overrides of IS? Second, are some of these factors more common in some situations than others? Third, how can the design of IS be revised in response to these factors?
5 Integration with Deception Reasoning
However, the same review shows that there is little research with the aim of integrating simulation and optimization in the solution-seeking process of complex and non-linear problems; only one article was found in this area. Several research efforts use agents and optimization and a few of them consider the case of a dynamic environment. Some of the researchers that use multi-agent architecture to solve optimization problems suggest the possibility of including simulation in the process. Also, there are some papers that consider agent-based modeling and simulation. Finally Baker suggests the use of agent-based simulation and artificial intelligence to solve optimization problems.
Following a review of these developments, it is clear that the integration of agent-based simulation and optimization is a research frontier. Such future projects will advance the field of optimization and will lead to new solution-seeking processes for many complex and real-world problems that cannot be solved with classical methods.
6 Results Determination and Display
The Adversarial Reasoning Module will show user-oriented actionable results, such as those featured in the following table:
|Results |User-Oriented Actionable Information |
|COAs Available for the Next Move|Blue, Red, and Green/White COAs |
|and Recommendations |Specific strategies such as Red’s most likely and most dangerous options (with similar information |
| |presented for Green/White if relevant to the situation), with assessed probabilities and outcomes |
| |The highest payoff, lowest risk option available to Blue clearly highlighted, with supporting, |
| |actionable information |
| |The ability to get detailed COA information based on user request |
|Multi-Move Pathways |Optimal pathway highlighted, with supporting, actionable information |
| |Move-to-move summary, with assessed probabilities and outcomes |
|Summary Scorecard |Integrated Payoff-Risk Results |
| |Statistics Associated with User and Higher-Level Objectives |
|Insights |Red and Green/White Patterns |
| |Observed Patterns |
| |Inferred Rules |
| |Possible Changes from Earlier Patterns and Rules |
[pic]
Figure 2. Time Series Visualization co-developed by lead PI (Hsu) at NCSA
3 Proposed Performance Metrics
Our proposed work is aimed at improving the state-of-the-art in continual dynamic replanning in the specific area of responding to unanticipated changes in the physical environment and state of an autonomous agent. We focus on the problem of mapping and localization using single agents and teams of agents. This provides a motivating application and a topic of focused synergy among three areas in which the principal investigator and co-PIs specialize, namely: lifelong, incremental reinforcement learning; robust team restructuring and dynamic allocation of responsibilities in teamwork; and metareasoning architectures for autonomous agents.
1 Operational Metrics
For brevity, we omit discussion of related replanning and general-purpose RL research in this preproposal. We note, however, that some previous work addresses architectures for learning in the presence of concept drift and contextual changes that are not directly observable. One of the earliest and best-known of these architectures is the classifier system, which is based upon rule-based representation of observations and context, online reinforcement learning by genetic and evolutionary computation, and a method for temporal credit assignment (the bucket brigade algorithm). Harries and Horn address the problem of tracking locally stable concepts[3] in systems that learn decision surfaces for classification and apply batch (offline) metalearning to discover these concepts. Widmer and Kubat specifically the problem of concept drift in time-ordered domains with continuous state values.
2 Technical Metrics: Decision Making and Predictive Validation
The novel contributions of our proposed work are that it:
• adapts the rule-based representation of the traditional classifier system to a temporal graphical model (DBN) that can represent continuous state and continuous time
• brings the decomposable structure of DBNs and our hierarchical architecture (DBN mixture model) to bear on domains that may be locally but not globally stable
• fits into a new metareasoning framework for one or more agents that replan based upon policies and quality functions learned using context-sensitive, online RL
Technical Approach
1 Survey of the Current State of the Art
1 Related Work
Parunak shows the industrial agent applications presented in the Workshop on Industrial Agents of the Center for Electronic Commerce. The projects fall naturally into four application areas: manufacturing scheduling, control, design collaboration, and agent simulation. There is not a project with an emphasis on optimization.
Davidsson, Johansson, Persson, and Wernstedt, mention that recently a number of agent-based approaches have been proposed to solve different types of resource allocation problems. The authors compare the strengths and weaknesses of Agent-Based approaches and classical optimization techniques and evaluate their appropriateness for a special class of resource allocation problems, namely dynamic distributed resource allocation.
The authors conclude that the properties of agent-based approaches and traditional optimization techniques complement each other. It would be nice if we could take the good properties from agent based approaches and combine them with the good properties of the mathematical optimization techniques. They plan to experimentally verify the conclusions of the theoretical analysis regarding the properties of the agent-based and classical optimization techniques. There was no discussion of agent-based simulation.
Wu, Yeung, Poon, and Yen prove that Internet and multi-agent systems may be a new alternative to planning for system expansion. The authors developed an Internet-based multi-agent system to help participants identify better or more appropriate partnerships (or coalitions) within the electricity market, such as, suppliers and consumers. Multi-agent systems are special type, which focus more on coordination and communication among agents to collaboratively accomplish a task. In this multi-agent system, the agents are naturally the generators, customers and coordinators. This research can be considered a kind of simulation but it is not related to optimization.
Parunak, Brueckner, Sauter, and Savit developed a system to solve the resource allocation problem, where each agent represents a different task consumer or supplier in allocation negotiations. The authors didn’t use simulation in their research.
Ouelhadj, Cowling and Petrovic proposed an agent architecture for integrated dynamic scheduling techniques of the continuous caster. Each of the process is represented by an agent, including the continuous caster agent, hot strip mill agent, the slabyard agent and the user agent. The authors show the scheduling problem of the continuous caster, it was modeled using a mixed integer-programming model and solved using constrained bin-packing heuristics and tabu search. In the multi-agent system the continuous caster agent generates predictive schedules using tabu search. For reactive scheduling, different measures were proposed to evaluate the effect of real-time events on the predictive schedule, and use these measures to make the decision whether to proceed with a schedule-repair or complete reschedule strategy to react to the real-time events. This research doesn’t consider simulation but it is developed in a dynamic environment where agents are permanently reacting and providing new solutions using Tabu search.
Loo, Lin, Kam and Varshney studied an architecture consisting of autonomous homogeneous agents, who are able to communicate with their peers. When the conditions allow communication, the agent would use this opportunity to modify its goals and plans, making use of this information exchanged with other agents. One of the potential advantages of communication among peers is that it may speed convergence. The model represents a group of cooperating vehicles (smart bombs, robots) that moves toward a set of (possibly moving) prioritized destinations. The objective is to maximize the number of encounters between the vehicles and the high-priority destinations. This research doesn’t talk about simulation but it is a case of agents and optimization in a dynamic environment.
Clements, Crawford, Joslin, Nemhauser, Puttlitz, Savelsbergh proposed a hybrid architecture, integer programming and heuristic search techniques, on a scheduling problem that arises in fiber-optic cable manufacturing. The proposed architecture of their local search algorithm has three components:
• Prioritizer. Generates a sequence of jobs, with higher “priority” jobs being earlier in the sequence. Uses feedback from the Analyzer to modify previously generated sequences.
• Constructor. Given a sequence of jobs, constructs a schedule. Uses “greedy” scheduling for each job, in the order they occur in the sequence, without backtracking.
• Analyzer. Given a schedule, analyses that schedule to find the “trouble spots”. This feedback is provided to the Prioritizer.
They use random moves to avoid become trapped in local optima. This architecture is a general framework, and not itself a specific algorithm. This research can be classified as a case of agents and optimization, but the simulation aspect is absent.
Bruzzone, Orsoni, Mosca, and Revetria, proposed an artificial intelligence (AI) based technique for optimization of the fleet management in maritime logistics. The optimization module uses AI techniques based on Genetic Algorithms (GAs) to suggest specific values for the case-dependant parameters, which customize the algorithms and criteria built into the decision heuristics to suit the particular situation (problem) at hand. Based on such customization, the logical flow of the heuristic decision steps automatically generates a tentative configuration of transportation routes and resource allocation, which can be tested for actual effectiveness in the process simulation environment. The outcome of the simulation test is fed back to the GAs where it serves as a guideline for the generation of improved sets of customization parameters. The same procedure is iterated until the optimal logistic configuration is found. During this iterative process, in the search of the optimum solution, the impact of each logistic configuration can be tracked over time in the dynamic simulation environment to assess their long, medium, and short term impact on performance.
While several approaches can be pursued for the optimization of the entire product transportation network (e.g. Artificial Intelligence techniques, Linear Programming, Simulation, etc.) none of them can address the whole problem alone.
The major advantage of using GAs in the optimization process is that the search for the optimum solution begins from an entire “population” of scenarios and, thus, from multiple points in the space of the possible solutions, which highly increases the chances of finding the actual optimum, rather than a sub-optimum.
The interactive use of AI techniques and simulation constitutes a hybrid approach, which the authors have extensively applied to address complex decision making issues in supply chain management. The approach has successfully supported the development of decision support systems typically combining either artificial neural networks or genetic algorithms with simulation in a variety of industrial context. This research considers AI, simulation and optimization, but the concept of the agent is not examined.
Baker, describes an agent as basically a self-directed software object. There are three types of architectures which are commonly studied in the multi-agent systems research community: functional, blackboard, hierarchical architectures, and heterarchical architecture.
• In a functional architecture each agent represents a functional capability. Usually there is only one agent per function.
• In a blackboard architecture, each agent has expertise in a certain area and the agents share their expertise by posting partial solutions to a problem on a central blackboard.
• In a hierarchical architecture, there are multiple levels of master/slave agent relationships where agents at one level of the hierarchy are slaves to a master agent at the next highest level of the hierarchy.
• In a heterarchical architecture agents communicate as peers, there are not fixed master/slave relationships, each type of agent is usually replicated many times, and global information is eliminated. Some advantages of heterarchical architectures include self-configuration, scalability, fault-tolerance, reduced complexity, increased flexibility, and reduced costs. The heterarchical architecture also enables massive parallelism.
He describes the factory control algorithms that can be put in a multi-agent heterarchy:
1) Dispatching Algorithms
2) Scheduling Algorithms
a) Optimal Scheduling
i) Simple Optimal Scheduling Problems
ii) Combinatorics
iii) Mathematical Programming
b) Nearly Optimal Scheduling
c) Towards Optimal Scheduling
i) Simulated Annealing
ii) Genetic Algorithms
iii) Neuro-scheduling
3) Heuristic Scheduling Methods
a) Forward/Backward Scheduling
b) Deterministic Simulation
c) Intelligent Scheduling
4) Pull Algorithms
Baker claims that an active area of research is how to make deterministic simulation algorithms and near-optimal (Lagrangian relaxation) algorithms fit the heterarchical agent architecture. Little research has been done with fitting simulated annealing, or genetic algorithms in this architecture, though successful uses of these algorithms in centralized systems have been reported. Though agent-based scheduling may be considered a subfield of AI-based scheduling, little work has been reported on making some of the standard AI-scheduling techniques work in a multi-agent heterarchy, though use of blackboards and functionally heterogeneous agent architectures is normal in that community. Baker dismisses the option of using agent-based simulation and artificial intelligence to solve optimization problems, claiming that there is little research in this field.
Rebollo, Julian, Carrascosa, and Botti apply a multi-agent system to the whole container allocation process in which it is desirable to minimize the containership stowage time. Container terminal management is a very complex system; it may be that the only way it can reasonably be addressed is to develop a number of modular components that are specialized for solving a particular aspect of it. To design the system architecture, the system has been divided according to its main tasks. Therefore, a different kind of agent for each one of the main tasks to be done has been modeled. These agents are mainly characterized by their independence from the rest of the system elements. They are able to coordinate and to communicate some decisions to the rest of the system. This architecture, which is designed according to the multi-agent paradigm, allows them to divide the problem into subproblems. Each subproblem can be solved by a specific agent. The proposed distributed approach enhances flexibility, efficiency and robustness. The agents use heuristics to solve the problem. Some of them use non-supervised learning techniques. A first prototype was being developed when the paper was written. This research includes agent modeling and optimization, but the simulation process is neglected.
Many of the studied articles refer to Glover’s Tabu search. Good heuristic procedures are based on ideas that can trace their origins equally to the fields of artificial intelligence and operations research. Today we can see several research efforts that lead to take advantage of these two areas of knowledge. It is clear that there is an increasing number of artificial intelligence techniques, including agent architecture, applied to the solution of complex and non-linear optimization problems
2 Relational Data Mining (NSF-Funded Effort, 2002-present)
[pic]
Figure 3. Prototype relational data mining system under development by proposer (KSU-CIS).
This section motivates the research and development plan for the technical objectives listed in section B and discusses our specific online RL methodology.
The focus of this research program is on formalizing data mining problems in recommendation of dataflow and COA components and representing them using graphical models of probability such as Bayesian networks and relational and object-oriented extensions thereof. A motivating application is to help users design new computational experiments in computational science and engineering by reusing software, data, and COA models produced in previously recorded experiments. Our central hypothesis is that probabilistic learning and reasoning are effective tools for building decision support systems and that relational graphical models learned from data can increase its scalability and efficiency. The primary contribution of this research shall be the novel combination of statistical algorithms for learning the structure of graphical models from data with existing techniques for constructing relational models of COA specifications, other repository contents, and user histories. The desired outcome of applying our new technology is the first content-based decision support system for scientific computing based upon statistical learning of dynamic relational graphical models. The technical objectives center around development of new algorithms for content-based CR using relational graphical models, statistical query optimization methods to make CR more efficient and scalable, and validation experiments to evaluate these on an application test bed in computational science portals.
3 Real-Time Decision Support (ONR-Funded Effort, 2000-2002)
We have formulated several probabilistic inference problems pertaining to dynamic, interactive decision support in personnel science. Specifically, we have developed probabilistic inference specifications for decision support problems in the following areas:
- Selection
o Problem definition: as for assignment, with different (more general) observables and criteria
o Observables: as for assignment, but with less historical evaluation data
o Unknowns: probabilistic ranking measures for selection
o Approaches: BN structure learning using K2, sampling-based approximate inference; comparison with fuzzy outranking
- Classification
o Problem definition: as for selection and assignment, with additional observables and more specific criteria
o Observables: as above; in addition: task-specific aptitude tests, qualifications, qualitative interview data, professional credentials, work history
o Unknowns: probabilistic ranking measures for selection
o Approaches: BN structure learning; sampling-based approximate inference; comparison with exact Bayesian network inference
To achieve the scientific goals of the research, we developed a Bayesian network inference system in Java using an XML-based infrastructure built upon the XML Bayesian network interchange format (XBN). This infrastructure is based in part upon technology developed by the PIs in collaboration with the National Center for Supercomputing Applications (NCSA). It consists of a software framework for rapid prototyping of distributed applications of knowledge discovery in databases (KDD). Basic research contributions from this second prototype are reported as follows.
[pic]
Figure 4. Screenshot from Bayesian Network tools in Java ().
The ultimate goal is to use this information to generate decisions using techniques such as:
- Bayesian network inference
o Deterministic: Lauritzen-Spiegelhalter
o Approximate: multiple stochastic sampling-based algorithms
- Multi-attribute decision making (e.g., value iteration)
- Multi-objective decision making (e.g., Pareto optimization)
Preliminary research (supported by ONR grant N00014-00-1-0769), we found that like many decision support problems, the above selection and assignment problems have ill-defined independent variables. That is, relevant variables are not fully identified. Traditional approaches to this problem include principal components analysis, independent components analysis, factor analysis, and other dimensionality-reducing transforms (e.g., self-organizing maps). Another approach uses wrappers for supervised inductive machine learning.
Based upon the above requirements, we have developed a suite of reusable Java-based software modules for reasoning under uncertainty in the personnel science domain – specifically, outranking measures for selection and predictive models for assignment and distribution. The results of this distributed data mining application will be compared to known historical observations and Bayesian optimal classification and prediction over these observations.
We have focused on development of input, data preparation (sampling, aggregation, and online analytical processing), structure learning, and sampling-based inference, especially adaptive importance sampling, and visualization and evaluation modules for constructing graphical models of probability – specifically, discrete Bayesian networks (BNs). We have developed modules in Java that implement K2, a BN structure learning algorithm, a gradient-based parameter estimation algorithm for learning conditional probability tables from data, and five stochastic importance sampling algorithms:
1. Forward simulation
2. Probabilistic logic sampling (rejection sampling)
3. Backward sampling
4. Heuristic importance sampling and self importance sampling
5. Adaptive importance sampling
6. Genetic algorithm for inference
1. Optimized implementation of K2
2. Genetic algorithm for structure learning
3. Probabilistic relational model with structure learning
A genetic wrapper has been shown to provide improvement in variable ordering when used in tandem with our (sampling-based) inference and structure learning algorithm (K2). We further hypothesize that it can be used to improve the performance of other extant structure learning algorithms. Our implementations of exact inference and adaptive importance sampling have been shown to be competitive with other inference algorithms in efficiency and accuracy, with the added benefit of being very robust in the presence of unlikely evidence. Preliminary studies of the Enlisted Master File completed since the conclusion of this project indicate that this problem environment is typical. Informally, we hypothesize that fast inference will provide a crucial functionality in interactive decision support – more precisely, soft real-time (smooth utility function) requirements.
SIGNIFICANCE: Our studies have provided information on to how an integrated learning and Bayesian network inference system can be applied to real-time, interactive, dynamic decision support in the domain of personnel management.
4 Simulation-Based Optimization (ONR-Funded Effort, 1999-2002)
Simulation is a powerful tool used by decision-makers to improve business and industrial operations. The idea is to simulate, using a model, a physical process on the computer, incorporating the uncertainties that are inherent in all real systems. The model is “run” to determine the consequences of the process. Then an analysis can be performed whose results can help to make decisions.
Multi-agent based simulation differs from other kinds of computer-based simulation in that (some of) the simulated entities are modeled and implemented in terms of agents, Davidsson. Davidsson compares multi-agent based simulation (MABS) versus discrete event simulation (DES). In MABS there is a close match between the entities of the reality, the entities of the model, and the entities of the simulation software. He argues that MABS has the following important advantages compared to more traditional DES techniques:
- MABS supports modeling and implementation of pro-active behavior, the agents are able to take initiatives and act without external stimuli.
- MABS supports distributed computation in a very natural way.
- Since each agent typically is implemented as a separate process and is able to communicate with any other agent using a common language, it is possible to add or remove agents during a simulation without interruption. This enables extremely dynamical simulation scenarios.
- It is possible to program the simulation model and software on a very high level, making it easier for non-programmers to understand and even participate in the software development process.
A disadvantage is that MABS compared to DES uses more resources, both for computation and communication, which may lead to slower simulations.
Multi-agent based simulation is nowadays used in a growing number of areas, where it progressively replaces the various micro-simulation, object-oriented or individual-based simulation techniques previously used. It is due, for the most part, to its ability to cope with very different models of “individuals”, ranging from simple entities (usually called “reactive” agents) to more complex ones (“cognitive” agents). The ease with which modelers can also handle different levels of representation (e.g., “individuals” and “groups”, for instance) within a unified conceptual framework is also particularly appreciated, with respect, for instance, to cellular automata. This versatility makes MABS emerge as the support of choice for the simulation of complex systems.
Contrary to numerical simulation, where their knowledge can only be represented by variables and relationships between variables, MABS allows, in theory, for a much wider range of representations: formulae, rules, heuristics, procedures, etc.
Dogoul, Vanbergue and Meurisse present a diagram to serve as guidelines for designers of multi-agent simulations, or evaluation framework for existing MABS.
The focus of agent based simulation on complex and non-linear problems is analyzed by Hans. He states that over time, two major non-linear modeling techniques emerged, agent-based and system dynamics modeling.
The aim of agent-based (or individual-based) modeling is to look at global consequences of individual or local interactions in a given space. Agents are seen as the generators of emergent behavior in that space. Interacting agents, though driven by only a small set of rules which govern their individual behavior, account for complex system behavior whose emergent dynamic properties cannot be explained by analyzing its component parts. Emergence, thus, is understood as the property of complex systems where “much (is) coming from little”. Emergence is the focal point of what now is called the theory of Complexity.
Simulations to analyze complex systems are characterized by both large parameter spaces and nonlinear interactions. Unfortunately, these same characteristics make understanding such models using traditional testing techniques extremely difficult. Miller shows how a model’s structure and robustness can be validated via a simple, automatic, nonlinear search algorithm designed to actively “break” the model’s implications. Using active nonlinear tests (ANTs), a simulation’s structure can be probed for key weaknesses and thereby begin to improve and refine its design.
Based on the scoring system, the ARM will be able to identify optimization strategies, both for a given move at a specific point in time and along a pathway of moves that takes place over an extended duration. The optimization strategies will specifically address Red Strategies, including Red’s most likely option (assessed both from Red’s perspective and from Blue’s view of Red) and Red’s most dangerous option (again assessed from Red’s perspective and from Blue’s view). Because Green/White interactions are important considerations, scoring will take into account effects on or from Green/White (to include the identification and assessment of specific Green/White strategies where appropriate). The ARM will also present optimal Blue strategies, indicating the options that offer the highest payoff, lowest risk combinations.
Because results will be probabilistic rather than deterministic, Monte Carlo statistics will show assessed likelihoods of the possible outcomes associated with a specific move or for various pathways associated with multiple moves that account for different branches and sequels.
5 Performance-Based Measures of Skill in Microworld Simulations (NSF, FAA & ONR-Funded Effort, 1996-1998)
[pic]
Figure 5. Performance-Based Measures of Skill for Range Fire Control
6 Immersive Training Systems for Crisis Management (ONR-Funded Effort, 1995-1998)
[pic]
The lead PI has over 9 years of experience in developing software tools for simulation, scenario generation, and visualization in crisis management, including DC-Train, a system developed at the University of Illinois for shipboard damage control. This system featured:
• simulation-based immersive training – automatic generation of scenarios to primary damage (mine/missile hit, shipboard fire) specification: Grois, Hsu, Voloshin, Wilkins (1998) used noisy-OR Bayesian networks to model scenarios and causal dependencies
• intelligent critiquing – identification of chief causes of kill points and review with (damage control assistant) student trainee
• learning and inference for predictive validation – temporal artificial neural networks and graphical models
• situational awareness – automated reasoning for recommended course of action
2 Approach for Technology Development
The Adversarial Reasoning Module will identify potential actions – encompassing the full range of available methods, means, and targets – and will assess and highlight the best options available to Blue.
Automating the development and analysis of COAs will involve work that addresses the actual situation, Blue perceptions of the situation as well as the Blue assessment of Red’s perceptions, and Red perceptions of the situation as well as the Red assessment of Blue’s perceptions. Outcome calculations will be based on ground truth, but decisions will be made based on Blue and Red perceptions (with these perceptions being modified over time by the outcome feedback each is able to obtain).
1 Data Modeling: Ontology and Representation
High-impact functions of this recommender system that are beyond current technology include the ability to:
• identify relevant features of terrain and scenarios that make it useful or a risk in specific ways
• recommend a COA to a user (CO) based upon queries related to objectives, available information
• adapt or repurpose COAs retrieved from a database;
• tell users how one COA may be adapted to suit a new specified purpose
• explain similarities between two simulation outcomes that make one informative to another
• broaden search selectively across data sources and transformation steps
• select COA components based upon context of a partially constructed COA or plan
• rate COAs by functional measures, not just usage statistics.
2 Algorithms for Automated Learning and Reasoning
We pose the questions:
• Are there class-level feature selection and construction techniques that are more resistant than existing ones to autocorrelative bias and linkage-based bias in learning RGMs?
• Can the decomposition approach discussed above be used to factor RGMs of COAs into modular groups of tables?
• What are the limitations of probabilistic relational models having to be directed acyclic graphs? Is it instead appropriate to model our data using relational extensions of undirected models such as Markov random fields (MRFs) or temporal models such as dynamic Bayesian networks (DBNs)?
3 Key Ideas for Future Development
We propose to investigate the problem of reinforcement learning with hidden context and concept drift. We will develop modules in our test bed to simulate mapping and localization by single agents and coordinated teams of agents, allowing a scenario designer to specify hidden contextual parameters: terrain type in a synthetic map, punctuated or incremental failures in sensors, communications, or whole agents in a multi-agent team, and damage to motive components such as treads and wheels. The simulator will be used for reinforcement learning experiments, facilitating our development of algorithms for online, active RL given prespecified context-modifying events. We will then instrument these RL algorithms on real agents and experiment with inducible hidden changes in context such as variable terrain severity and localization breakdown in other team members. We will develop new algorithms for learning the structure of dynamic graphical models of probability that adapt context-specific independence to time-ordered domains by learning: (1) time lag parameters; (2) hidden variables that correspond to specific types of context, such as bias and other problems in sensors, teammates, terrain severity, and shared responsibilities; and (3) causal dependencies between these hidden variables and directly observable variables such as sensor readings.
4 Self-Evaluation Methodology
We believe the significant time investment that the ARM represents is worth while because it provides several benefits towards validation of our solution approach. First, it gives us the chance to experiment with real data, which in data mining tends to reveal practical issues, such as biases and sparsity in data acquisition for recommender systems, that are hard to simulate. Second, the motivation that it gives us to consult with biologists strengthens our interdisciplinary work and helps us to better understand and communicate its broader implications. We will evaluate our system using several approaches: first by measuring the accuracy and robustness of learned RGMs and second by assessing the quality of recommendations.
Ablation Studies: Our hypothesis that relational graphical models are the appropriate representation and probabilistic reasoning the appropriate techniques for building the classification, prediction, and reasoning models for CR shall be tested by comparing the level of reuse (percent of new code implemented and development time) without intelligent ARM, with the current ARM technology, and with the ARM system at reduced levels of accuracy for learning the models described in previous sections.
Accuracy and Reliability Measures: We will look at measures adopted from information extraction that also serve for COA retrieval, namely precision and recall of RGMs. Another absolute measure of quality is log likelihood of models with respect to data; where appropriate (i.e., where norms can be established) and necessary (i.e., ground truth is not available), we will consider such measures. We will also evaluate robustness of individual relations and memberships under cross-validation and bootstrap analysis. Finally, we will look at prediction accuracy over rankings and ratings; these may be weighted using regression or according to a utility or prioritization scheme cf. information retrieval. Where applicable, we will also compare our attribute selection and construction in flat structure learning to algorithms such as TETRAD and the Sparse Candidate algorithm.
Gain from Recommendations Made: The motivating goal of the ARM is to benefit the warfighter in design and implementation of computational experiments. Consulting with some of these bioinformaticians such as our current collaborators affords us the chance to elicit ground truth (gold standard RGMs), but this is generally feasible only for static relational attributes. Dynamic relational attributes are known to be difficult to elicit, especially for complex domains. Nevertheless, we can evaluate CR quality by looking at the distribution of posterior ratings (after COA use) versus preliminary ratings and recommendations. The software engineering literature provides metrics for reuse and analytical and empirical evaluation methodology that we shall apply.
Solution Quality Assessment in Computational Sciences: Focusing on experiment design in computational genomics, we will apply COAs developed and modified by the team members. Evaluations shall be carried out in cooperation with end users of the CR system in our collaborators' groups, to obtain more detailed assessments of solution quality than our user ratings.
In the ARM, we shall evaluate probabilistic representations in comparison with others such as non-probabilistic relational models developed using existing use-case and entity-relational modeling approaches, as well as against non-relational classifiers trained using flat data (WEKA) provides a large suite of popular inducers for such comparative experiments, as does our toolkit, Bayesian Network tools in Java (BNJ).
Our research program has two major aims: to advance the understanding of computational methods for data mining in the domain of recommender systems for COA components, especially in Grid computing, and to apply these methods to improve on existing COA development practices. Therefore, we will evaluate progress toward these two goals in different ways using criteria appropriate to each.
Management Approach
1 Statement of Work
American Systems Corporation’s Joint and International Weapons & Concepts Directorate has tremendous operational and analytical expertise that is directly relevant to the development of the proposed Adversarial Reasoning Module. Drawing on retired military officers and experienced analysts – with time on the ground in Vietnam, Korea, Somalia, Haiti, Bosnia-Herzegovina, Kuwait and Iraq – ASC will bring to bear subject matter expertise in Urban Operations, small-unit capabilities and tactics, Red Cell assessments, Course of Action development and assessment, Command and Control and Decision-making, plus “grey beard” consultation.[4] This expertise will be applied in the following areas:
Predictive Analysis
– Characterizing the Urban Environment
o Identifying factors – whether from the complex, man-made physical terrain, the diverse population, or the physical or service infrastructure – that affect decision making and/or outcomes
o Analyzing urban effects such as reduced Intelligence, Surveillance, & Reconnaissance (ISR) capabilities; degraded communications; slower and more decentralized Command and Control; 3-D maneuver through airspace, surface, and sub-surface corridors; interior and exterior movement; constraints on direct and indirect fires; and so forth
– Automating Course of Action (COA) development, analysis, and selection
o Identifying doctrine, tactics, the current situation and Red/Blue perceptions, and other factors that affect available options
o Generating distinct Course of Action (CoA) based on key factors
o Creating an operationally relevant means of scoring alternatives that supports automated COA analysis
o Selecting particular COAs of interest – such as most dangerous, most likely, and high-Payoff/low-Risk
– Determining/displaying results tailored to the Military Commander’s needs
o Developing ways to summarize diverse COAs to reduce the probability of surprise
o Determining how to identify and highlight COAs and pathways that lead to high-Payoff/low-Risk outcomes
o Integrating payoff and risk information to support real-time cost-benefit decisions
ARM Data Stores and ARM Metrics
– Assisting in the design and the development of the data stores for the Red, Blue, and Green Models; the Terrain Model; and the Common Operating Picture database
– Assisting in the design and development of the operational metrics to include the definition and self-evaluation approaches to be used in in-house experiments and other evaluation events.
2 Program Schedule
The following table depicts the key aspects of ARM development and expectations by the end of Phase I.
|Aspect |Expectations |
|Situation Awareness |Based on Ground Truth |
|COA Development |Rule-Based Course of Actions |
|COA Analysis and Selection |Single-Plane (Only Consider Actual Situation) |
|Results Determination/Display |Basic Scorecard |
The following table depicts the key aspects of ARM development and expectations by the end of Phase II.
|Aspect |Expectations |
|Situation Awareness |Perception-Based |
|COA Development |Rule-Based + Feedback |
|COA Analysis and Selection |Multi-Plane (Actual Situation + Players Perceptions) |
|Results Determination/Display |Robust Scorecard |
The following table depicts the key aspects of ARM development and expectations by the end of Phase III.
|Aspect |Expectations |
|Situation Awareness |Perception-Based and Highly Dynamic |
|COA Development |Adaptive, Based on Evolving Rule Set and Feedback |
|COA Analysis and Selection |Multi-Plane/In Depth (Can Look Many Moves Forward) |
|Results Determination/Display |Dashboard |
Finally, Figure 6 gives the overall project timeline, including research tasks, preliminary work, student employment, software, and workshops.
[pic][pic]
Figure 6. Master Project Timeline for ARM Development, Integration, Experimentation
3 Deliverables
Phase I Research: Information Gathering. In the initial phase of research, we propose to gather information about the extent of the IS override phenomenon. Our plan is to visit experts who work with advanced guidance systems at NASA, FAA, and USAF. We have contacts at each agency who have agreed informally to support our efforts.
We will start with 3-5 person focus groups by asking experts to describe situations where they either turned-off or override the output of IS. The goal will be to explore characteristics of situations that lead to these decisions. At first, we will offer no direction or guidance into the sort of factors we are considering. Only after the experts have engaged in a spontaneous discussion will we explore the 11 psychological factors identified above. This 2-stage approach was applied successfully to a study of auditing expertise recently conducted by M. Abdolmohammadi, J. Shanteau, and G. Searfoss.
The next stage will involve one-on-one interviews with these experts to explore the insights gained from the focus groups. For instance, we will ask for priority orderings of the factors mentioned in the focus groups, as well as the 11 factors identified above. We will also check for omissions and redefinitions of items. This will allow us to refine the list and to establish groupings of similar items. Such an approach was used in the PI’s previous research with expert Air Traffic Controllers at the FAA.
he goal from this phase will be to categorize a small number of reasons for the turn-off, override decisions, eg, concerns about the reliability of the systems or social/cultural factors. The definitions of these categories will be refined and verified by successive interviews with experts.
Phase II Research: Simulation Research. In the second phase, we will conduct experimental research with trained human operators working with specifically designed IS simulations. These tools will be based on modifications of existing guidance system software; this software is presently available at KSU for both robotic systems and air traffic control. Based on previous research conducted for FAA and ONR, we expect that participants can be trained to mastery in 3-4 weeks.
Once proficient, users will make operational decisions in a variety of scenarios. Built into each scenario will be instances of one or more of the factors identified in Phase I. For instance, the system may go down inexplicably (crash) without warning. Alternatively, the system may have delayed feedback built in. The users will be allowed either (1) to continue using the system as is, (2) to override all or part of the system output, or (3) to ignore the system altogether.
In addition to the choice of whether (and how) to use the system, overall performance will be evaluated using techniques described in a forthcoming paper (in Human Factors) by D. Weiss and J. Shanteau. These techniques are based on applications of a novel approach to performance assessment that does not require unique identification of a gold standard.
This empirical research will allow us to test hypotheses related to use or non-use of IS in guidance systems. Based on Phase I research, for example, we might predict that attention overload leads to a higher rate of system turnoff than lack of control. We will also look for interactions where multiple factors are operating in a scenario, eg, an interaction between system crashes and mode mismatch problems.
Phase III Research: Design Recommendations. Based on the results from Phases I and II, we will explore directions for improving the design and implementation of IS. These recommendations will suggest alterations in standard best practices in software/hardware design. Based on input from the psychologists on this project, the computer scientists on our team will lead this phase of work.
Several directions for design recommendations are possible. If the results of Phases I and II suggest that more transparent identification of operating modes is needed, then improved methods for displaying system states would be recommended. Similarly, if miscalibration of inputs is a concern, then routines for encouraging accurate calibration would be incorporated (D. Hilton).
These design modifications will be checked first by incorporating them into the simulations routines used in Phase II. Thus, we will compare performance with and without the incorporation of the design changes into our simulated guidance systems.
We will verify the design recommendations with the experts used in Phase I. We will ask them about the reasonableness, as well as implementation potential, of each recommendation. Where appropriate, we will modify our recommendations to reflect the feedback obtained from these experts.
4 Cost Summary
The following tables (explained in detail in the cost volume) detail the cost proposal for the three proposing organizations. The requested cost vehicle is a grant.
Table 1. Cost Proposal Summary (Table 5 from PIP)
|COST ELEMENT |GFY 05 |GFY 06 |GFY 07 |Total |
| | | | | |
|Technical Labor |$348,557.40 |$357,353 |$366,589 |$1,072,499 |
| | | | | |
|Travel |$37,000 |$37,000 |$37,000 |$111,000 |
| | | | | |
|Subcontractors | | | | |
|(1) American Systems Corporation |$411,155 |$560,400 |$575,256 |$1,546,811 |
|(2) University of Mississippi |$382,102 |$388,736 |$403,905 |$1,174,743 |
| | | | | |
|E. Other Direct Costs |$19,930 |$3,930 |$19,930 |$43,790 |
| | | | | |
|F. Total Direct Costs (A through E) |$1,198,744.44 |$1,347,419 |$1,402,680 |$3,948,843 |
| | | | | |
|I. Indirect Costs |$209,524 |$183,210 |$194,819 |$587,553 |
| | | | | |
|J. Total Direct and Indirect Costs (H + I) |$1,408,269 |$1,530,629 |$1,597,498 |$4,536,396 |
Table 2. Budget, by Contracting Organization (Table 5 from PIP)
|Organization |GFY 05 |GFY 06 |GFY 07 |
|Prime (KSU) |$615,012 |$581,493 |$618,337 |
|University of Mississippi |$382,103 |$388,737 |$403,903 |
|American Systems Corporation |$411,155 |$560,401 |$575,258 |
|TOTAL |$1,408,270 |$1,530,629 |$1,597,498 |
Table 3. Proposer's Team Budget, Year 1 ( Table 8 from PIP)
| |GFY 05 |
| | |
|BASE |OCT |
|BASE |OCT |
|BASE |OCT |
|BASE | |
|A. Direct Labor - Dollars |$1,072,499 |
| | |
|B. Direct Labor - Hours | |
| | |
|C. Travel |$111,000 |
| | |
|D. Subcontractors | |
| (1) American Systems Corporation |$1,546,811 |
| (2) University of Mississippi |$1,174,743 |
| | |
|E. Other Direct Costs |$43,790 |
| |$0 |
|F. Total Direct Costs (A through E) |$3,948,843 |
| | |
|G. Indirect Costs |$587,553 |
| | |
|H. Total Direct and Indirect Costs (F + G) |$4,536,396 |
5 Personnel
Kansas State University
Dr. William H. Hsu, Ph.D.
Professional Preparation
University of Illinois at Urbana-Champaign (UIUC), Computer Science, Ph.D., 1998
The Johns Hopkins University (JHU), Computer Science, M.S.E. (Master of Science in Eng.), 1993
The Johns Hopkins University (JHU), Comp. Sci. / Math. Sci., B.S. (first in class / honors), 1993
Appointments
Assistant Professor Department of CIS, Kansas State University, 1999-present
Research Scientist National Center for Supercomputing Applications, 1998-2003
Honors and Awards
• NSF EPSCoR First Award, 2002
• NCSA Industrial Grand Challenge Award, 1999
Synergistic Activities
• Conference and workshop committees: Co-chair, AAAI/KDD/UAI-2002 workshop W18 on Real-Time Decision Support
• High-performance computing and software: lead developer, KSU Bayesian Network Tools in Java (2000 - present); lead developer, KSU Machine Learning in Java (2000 – present); NCSA Data to Knowledge (D2K) development team (1998 – 2002)
• Editorial boards: Intelligent Data Analysis (2003 – present) and journal paper refereeing: Info. Sciences (special issue on Soft Computing and Data Mining, 2002), Machine Learning (2001 – present), J. Machine Learning Res. (2002 – present), IEEE TKDE (1997)
• Student learning and professional development facilitator: faculty advisor for Computing Research Association (CRA) Collaborative Research Experience for Women (2002 – present); KSU-CIS undergraduate honors program chair (2001 – present); co-organizer of KSU Engineering and Science Summer Institute (2001, 2002)
Thesis Advising and Postgraduate Sponsorship
Haipeng Guo (PhD 2003): Hong Kong University of Science and Technology
Siddharth Chandak (MS 2003): practical training; Vinod Chandana (MS 2003: practical training; Raju Mantena (MS 2003): practical training; Benjamin B. Perry (MS 2003): Quantum Leap, DE, USA;
Sreerama N. Valluri (MS 2002): unknown; Afrand Agah (MS 2001): PhD student, University of Texas – Arlington, USA; Steven M. Gustafson (MS 2000): PhD candidate, University of Nottingham, UK
Summary of Graduate Students Supervised
To date, July 1999 – February 2004:
PhD: 1 graduated, 2 current; MS: 7 graduated, 2 current; M. Software Eng.: 4 graduated, 1 current
University of Mississippi
Dr. Norman Keith Womer, Ph.D.
Professional Preparation
BA, Economics, Miami University, 1966
PhD, Economics, The Pennsylvania State University, 1970
Appointments
- Director, Hearin Center for Enterprise Science University of Mississippi, 2001- present
- Interim Dean, School of Business Administration, University of Mississippi, 1999-2001
- Associate Dean, School of Business Administration, University of Mississippi, 1998-1999
- Chair, Department of Economics and Finance, University of Mississippi, 1986 – 1998
- Professor, Department of Economics, Clemson University, 1984-1986
- Professor, College of Commerce and Industry, Clemson University, 1979-1984
- Associate Professor, School of Engineering, Air Force Institute of Technology, 1973-1979
- Assistant Professor, Department of ORAS, Naval Postgraduate School, 1969-1973.
Synergistic Activities
- Editor, Topics in Operations Research
- Associate Editor, Military Operations Research
- Editor-At-Large, Interfaces (1995-2000)
- Board of Trustees, Wood College
- Chair, Mississippi Consortium for Military Personnel Research
Graduate Students Supervised
Thomas Gulledge (George Mason University)
Andy Litteral (University of Richmond)
Dennis Riddley (Florida A and M University)
Gary Fellers (Augusta College)
Jeff Camm (University of Cincinnati)
Bashir Al-Abedalla (University of Jordan)
Young Silk Kwak (Delaware State University)
Che-Peng Lin (Chiu Feng University)
American Systems Corporation
LIEUTENANT GENERAL EMIL R. (BUCK) BEDARD, USMC, RETIRED
A retired three-star General with tremendous command experience – at the Company, Battalion, Regiment, Division, and Marine Expeditionary Force (MEF) levels – General Bedard completed his career as the Chief Planning and Policy Officer of the U.S. Marine Corps. This includes first-hand urban operations experience while commanding forces in Mogadishu during Operation Restore Hope. In addition, he led the Marine Corps’ organizational and tactical planning in a time of unprecedented change; fulfilled training, staffing, and equipping responsibilities for Marines engaged in Afghanistan and Iraq; displaced Marine Corps Headquarters command and control during the 9/11 attack on the Pentagon; and deployed forces throughout the world in support of peacetime engagement and contingency operations. Instrumental in determining training requirements and measures of effectiveness, General Bedard played a central role in increasing combat readiness throughout the Marine Corps.
JOHN J. NELSON
Mr. Nelson possesses vast experience analyzing complex defense issues. He is currently leading a NATO assessment of non-lethal weapons (NLWs) and their potential effectiveness across the spectrum of military operations, including in urban environments. His previous work for ASC – using Futures Analysis and scenario planning methods – should directly benefit the COA development and assessment elements of the ARM.
Before joining ASC, Mr. Nelson worked for 15 years with the Center for Naval Analyses. During this time, he headed numerous major studies. These included the Marine Aviation Requirements Study, several studies analyzing the organization of major Marine Corps commands (including process and resource allocation analyses), work on Service and Joint planning processes, and on-the-ground operations analysis experience with the Combined Joint Task Force in Somalia during Operation Restore Hope and with the Implementation Force in Bosnia during Joint Endeavor. This last item included extensive work developing and analyzing IFOR’s contingency plans. In addition, during his assignment with the I Marine Expeditionary Force, Mr. Nelson played the lead role in developing and conducting conferences and exercises focused on Humanitarian Assistance Operations, Peacekeeping, Peace Enforcement, and Urban Operations.
GEORGE P. FENTON
George Fenton, Colonel USMC Retired, is a senior executive with extensive experience in managing large programs, program risk assessment, resource determination, problem resolution, and organizational and strategic planning. Completing, nearly 29 years of military service as an infantry officer, Colonel Fenton has had operational experience in urban operations in Mogadishu and Kismayo Somalia (1993); military planning for operations in Haiti (1993-94); and served as the former Director of the United States Department of Defense Joint Non-Lethal Weapons Directorate (1998-2002) as well as Chairman of the NATO sponsored study, Non-Lethal Measures of Effectiveness Study in 2001-02. Colonel Fenton is recognized as an international advocate in championing the importance and use of non-lethal capabilities in military urban operations and law enforcement. He is presently under contract to bridge the program efforts of the Joint Non-Lethal Weapons Program and the Joint Urban Operations Office.
TIMOTHY J. FOX
Mr. Fox is an accomplished leader and manager with demonstrated skills as a planner and researcher. He has extensive experience in conducting operational analysis and assessment, managing training and education programs, conducting personnel selection and management efforts, and overseeing research efforts. He has successfully completed the Defense Acquisition University’s, Acquisition 101 and Logistics 101 and currently provides onsite day-to-day support to the Concepts and Requirements Division, Joint Non-Lethal Weapons Directorate. Managed, edited and readied for publication the Kosovo Incident Case Study, a battle study examining non-lethal weapons employed in Kosovo for the Joint Non-Lethal Weapons Program. He was also a key member of the team that developed the US/UK Non-Lethal capabilities in Urban Operations Wargaming program.
Mr. Fox has a detailed understanding of joint and service doctrine from his experiences in the Fleet Marine Force and at the Marine Corps Warfighting Laboratory, Wargaming Division. He has designed, organized and conducted seminar wargames, focusing on future military requirements at joint, combined and naval levels examining and analyzing operational concepts and new technology implications at strategic, operational and tactical levels of warfare.
BOBBY STRAIGHT
Mr. Straight has a diverse and extensive background encompassing military intelligence operations, exercise development/execution, modeling and simulation, project management, and operational analysis. While at ASC, Mr. Straight has been involved in the Verification, Validation, and Accreditation of the Joint Conflict and Tactical Simulation for non-lethal weapon applications; the development and conduct of studies to acquire a baseline comparative analysis of fielded non-lethal weapons; the analysis of existing and developmental non-lethal weapons compared to joint needs for the identification of capability gaps; and the revision of the Multi-Service Tactics, Techniques, and Procedures for Non-Lethal Weapons. Prior to joining ASC, Mr. Straight served as Commander, Opposing Forces, at the USMC MAGTF Staff Training Program. Mr. Straight is a retired USMC officer.
TIMOTHY FERRIS
Possesses significant on-the-ground urban experience during joint and combined relief and stability operations (Provide Relief and Restore Hope) throughout Somalia. Combined, Joint and Marine Corps career expertise in command and staff action, including production, review and editing of top-level joint and Service staff operation, planning, and logistics products; and documents of national and international impact. Served as executive assistant to General Norman Schwarzkopf, USA, and aide de camp to General Joseph P. Hoar, USMC during two nonconsecutive tours of duty at United States Central Command. Principal staff experience at the MEU, MEB and MEF levels (the latter an 0-6 position) as well as all Marine Corps organizational levels from company through regiment and training command assignments.
KSU: POSTDOCTORAL RESEARCH FELLOW
This postdoctoral researcher will have a background in probabilistic reasoning and learning, and will be responsible for conducting experiments using Bayesian network codes and working with the PI and research programmers to develop and enhance them. The PI is currently working with potential candidates for this position both at Kansas State University and several other U.S. universities.
6 Related Experience
This proposal is responsive to a number of program objectives as outlined in the solicitation. For instance, we will investigate how IS changes the nature of human activities and introduces new risks. We will examine how human experts can be effectively integrated into crucial activities without requiring continual monitoring of autonomous systems. We will look at what design principles for individual agents best support mixed-initiative human-automation systems. A particular target for this research will be to examine how shared frames of references (can) be designed and implemented to overcome mismatches between the knowledge modeling in the intelligent software and human understanding of operational activities.
The proposed research addresses priorities in other areas of the Announcement as well. Thus, we will be “developing effective techniques for performing diagnosis for both discrete and continuous failures” (p. A-3). Finally, this proposal has relevance for “expanding, implementing, and testing algorithms that use causal knowledge to predict the effects of events or actions (p. A-6).
PI Qualifications. The research team is well suited to investigate the issues raised in this proposal. The lead PI is experienced in judgment and decision making research, and has coordinated recent research projects funded by the FAA, ONR, NIH, NSF, and DOD. Moreover, he is the former Program Director of the Decision, Risk, and Management Science (DRMS) at NSF. For examples of work products (eg, publications, workbooks, and computer routines) for previous funded efforts, see ksu.edu/psych/cws
The co-PI’s are experienced in the design and evaluation of intelligent systems, especially in robotics. The second PI has taught and conducted research on software engineering for 25 years, with an emphasis on robotics. He was instrumental in building the KSU robotics laboratory. The third PI teaches/researches artificial intelligence. He has worked with many training simulators and has published extensively in the machine learning area.
7 Facilities and Teaming
KSU presently has most of the software and hardware needed to conduct the proposed research. The primary computational facility available to the team is a Beowulf cluster in the KSU Department of Computing and Information Sciences (CIS). The lead PI (Hsu) has access to this system, which consists of over a dozen dual-processor (Intel Pentium III and IV Xeon) and quad-processor (Intel Opteron) nodes. These shall be used for rapid prototyping only. All experimentation with the prototype and fielded system shall be conducted on the standard (2.5+GHz Pentium) desktop PC specified by the DARPA RAID BAA.
In addition, KSU has access to C-TEAM, a simulation of ATC created by the FAA for initial evaluation of air traffic controllers. This routine has proved to be remarkably stable and productive in previous research conducted by the PI. Moreover, the FAA recently provided an open-source version of the software to facilitate on-site design modifications.
KSU has dedicated three laboratories to running C-TEAM. Each of these labs is capable of running 4 (or more) operators as a team, with a server operating in an adjacent room. One of the labs is based on collocation of operators, ie, they are in the same room together. The second lab is based on distributed decision making, ie, operators are physically separated from each other. The third lab allows for up to 12 operators to work on complex teamwork problems.
This simulation is ideally suited for the proposed research for three reasons: First, scenarios can be generated that vary in complexity, density, duration, etc; this is a vital feature that allows us to design scenarios for a wide variety of circumstances. Second, a number of performance measures are available, including some we added to the original program; the most notable is the addition of CWS (Cochran-Weiss-Shanteau), a measure of expert performance without using gold standards. Finally, CTEAM contains a replay option that allows for complete playback of operator behavior in real time; this option is important when conducting detailed followup analyses.
Because of their unique capabilities, ASC and KSU have determined that it will be of mutual benefit to form a Team to propose to Defense Advanced Research Projects Agency the Team’s uniquely and specially developed expertise as appropriate and necessary to meet the management and technical objectives of the client’s project for the purpose of being awarded a contract to perform the Project. KSU and ASC mutually agree as follows:
1. Client and Project.
The Client is identified as the Defense Advanced Research Projects Agency. The project is identified as Solicitation Number BAA04-16, Real-time Adversarial Intelligence and Decision-making (RAID). For the purposes of this proposal effort and eventual contract selection and award, KSU will be the lead (Prime) and ACS will be a Consultant to the Prime. The Prime has designated Dr. William Hsu as the overall Principal Investigator. ASC has designated George Fenton as its representative in the proposal effort.
2. Proposal Efforts
Each party will exert diligent effort to contribute to the marketing effort, including but not limited to participating technical and strategy meetings, and producing all or portions of the proposal and presentation documents. The Prime and Consultant shall cooperate fully to promote the mutual benefit of the Team developed as a part of this Agreement.
Each party shall furnish, for incorporation into the marketing effort, including all materials pertinent to the services assigned for that party. The Prime shall select the format, content and delivery method of all proposal documents. As necessary to prepare the marketing documents, the Consultant will deliver in a timely manner, written materials for its portions of the scope of services.
To the extent not prohibited by applicable law, each party shall bear all the costs, risks and liabilities incurred by it arising out of its obligations and efforts under this Agreement during the marketing effort. Neither party shall have any right to any reimbursement, payment or compensation of any kind from any party prior to the award of the Project to the Prime and execution of a subcontract between the Prime and Consultant.
3. Other Provisions
Nothing in this agreement prevents the consultant from participating in or leading one of the other contracts.
This agreement does not bind the parties to create any joint venture, partnership or formal business corporation of any kind.
This agreement may not be assigned or otherwise transferred by either party in whole or in part without the expressed prior consent of the other party.
This agreement shall remain in effect until terminated by any of the following:
• Upon 30 days notice by either party
• The client’s announcement that the project is canceled or that the project is awarded to another party
• If client awards the Project to the Prime, upon execution of a Prime-Consultant agreement.
Subcontracting Plan: Please see Appendix.
8 Security Plan
Although the proposed work and deliverables are at the unclassified level, the principal investigators are cognizant of the sensitivity of some of the RAID results, input (training) data, and software components such as the Integration framework and the Experimentation test harness.
Assuring the security and integrity of information systems poses major challenges to government, industry, and academia. The President’s Information Technology Advisory Committee (PITAC) reminds us that IT security is of vital importance to the national security, economic welfare, and private citizens of our nation and of its allies. Within the domain of information security, three key areas are of immediate priority to this proposing team: (1) safety-critical software and networks; (2) anomaly detection for prevention of network intrusion; and (3) digital forensic analysis to protect the privacy of users and the integrity of networks and data.
Safety criticality shall be protected by using up-to-date Secure Sockets Layer (SSL) libraries for all distributed components of the ARM. As the ARM shall run strictly on a local area network, typically on one dual-processor Pentium IV PC or a pair of uniprocessor PCs, this is an acceptable level of risk.
Commercial-off-the-shelf (COTS) network monitoring tools maintained by each participating institution shall be used. Virus scanners, firewalls, and appropriate restrictions shall also be installed. For example, inetd (for Telnet/FTP) is omitted in favor of SSH (secure shell).
9 Statement of Rights Claimed for Software Deliverables
NONE.
Self-Assessment According to Evaluation Factors
1 Technical Depth and Feasibility
• Understanding of the current and projected R&D on predictive analysis: very high. The team has experience with situational awareness, access to subject matter expertise urban doctrine, and experience with urban crisis management simulations (e.g., RoboRescue Urban).
• Understanding challenges: high. One of the subcontractors (ASC) has primary responsibility as a procurer/clearinghouse for relevant subject matter expertise.
• Soundness of the technology approach at the component and systems levels: very high. Each of the team member institutions has investigators who were responsible for developing optimization, learning, and automated reasoning modules and succeeded in delivering satisfactory prototypes.
• Potential for revolutionary advances in addressing the technical challenges: high. ASC’s role in this project is primarily as a development subcontractor/data integrator, although some of the presented algorithms are experimental and proprietary. KSU and Ole Miss have basic research responsibilities.
• Justification of design choices as compared to alternative techniques: moderate to high. A variety of methodologies (graphical models and other Bayesian representations; dynamic and relational models) are presented, the plan is to experiment and compare, then select the best algorithm for the task.
2 Consistency with RAID Program Concepts
• Consistency with the RAID system and program concepts: very high. The ARM program concept is strictly based upon the functional requirements for the urban combat simulator and the problem formulation given in the BAA.
• Depth and specificity of the proposed effort’s system and program concepts: very high. The maturity of previous work by individual team members and KSU/Ole Miss working in tandem assures a high probability of success.
• Precision and coverage of the proposed effort’s metrics: high.
• Plan for collaborating with other developers, integrator: moderate to high.
3 Cost realism
• Cost-benefit ratio: extremely high.
• Realism of cost levels: extremely high.
• Cost-effectiveness: extremely high.
• Procurement competitiveness: very high.
4 Personnel and Corporate Capabilities and Experience
• Qualifications/experience: extremely high
• Availability: very high
• Soundness of personnel assignments: very high
• Collaboration off-site: high
• Adequacy of proposed facilities: high
Appendix: Subcontracting Plan
Type of Plan
Individual Contract Plan - Individual Contract Plan, as used in this subpart, means a subcontracting plan that covers the entire contract period (including option periods), applies to a specific contract, and has goals that are based on the offeror's planned subcontracting in support of the specific contract, except that indirect costs incurred for common or joint purposes may be allocated on a prorated basis to the contract.
2. Goals
As an agency of the State of Kansas, Kansas State University will follow all state purchasing guidelines and all Federal guidelines referenced in the contract document. State guidelines encourage the use of small business concerns, small disadvantaged business concerns, and women-owned small business concerns.
The following steps will be taken in conjunction with the requirements of the prime contract :
• Such firms will be placed on solicitation lists.
• Assure such firms are solicited whenever they are potential sources.
• Divide total requirements, when economically feasible, into smaller tasks or quantities to permit maximum participation by such firms.
• Establish delivery schedules where the requirement permits, which encourage participation by such firms
• Use the services and assistance of the Small Business Administration, and the Minority Business Development Agency of the U.S. Department of Commerce, as appropriate.
• Require any subcontractors, if sub-subcontracts are to be let, to take the affirmative steps listed above.
In addition, we are committed to following stated goals in Public Law 95-507, The Amendment to the Small Business Act, Public Law 106-50, The Veterans Entrepreneurship and Small Business Development Act of 1999, Public Law 100-656, Business Opportunity Development Act of 1988, Public Law 100-180, Section 806, Requirements for Substantial Progress on Minority and Small Business Contract Awards, and Public Law 99-661, Section 1207, Contract Goal for Minorities.
A. The total estimated value of all planned subcontracting activities, i.e., with all types of business concerns under this contract, will total $2,721,556. The total estimated value of planned subtracting for small business concerns will total $ 0.
The Work proposed under this solicitation is highly technical and will focus the talents among real-time intelligent systems, optimization, and judgment and decision-making researchers to develop an adversarial reasoning module (ARM) for DARPA’s real-time adversarial intelligence and decision making (RAID) requirement in the domain of urban combat. Thus, the allocation of the budget is heavily weighted toward the payment of salary, wages and fringe benefits for personnel of each Project Team Member. The proposed Project Team has been assembled based on the proven subject matter expertise of each of the respective parties. The interactive skill sets of the Project Team that will be exercised to perform this project is based on previous collaborative interactions amongst the key personnel, which results in the establishment of a complimentary cadre of experts. The planned subcontracting activities for this project will consist of the issuance of subcontracts to American Systems Corporation (ASC) and to The University of Mississippi by the Prime Contractor, Kansas State University. The subcontracted tasks are highly specialized and consist of the completion of certain specific technical objectives that have been identified, extrapolated and allocated to each Project Team Partner based on each partner’s specific intellectual and technical strength.
Due to the specific task to be completed under this solicitation and the expertise required, coupled with the previous collaborative efforts and the fact that the proposed subcontractors participated and were instrumental in the preparation of the proposal submitted by Kansas State University pursuant to this solicitation, the proposed subcontracts would be considered sole source. Based on the technical needs of the project objectives and the required funding, Kansas State University does not anticipate that any other subcontracting activities would be necessary.
However, given Kansas State University’s commitment to this procurement statute, it is anticipated that opportunities will exist for small businesses covered by this statute to be afforded the opportunity to participate in procurement opportunities relating to the purchase of certain goods and services, including material and supplies, certain travel expenses as well as certain miscellaneous other expenses included with in the proposed budget. KSU will exercise due diligence and will establish goals for these areas of expenditure for each reporting category including, Small Business Concerns, which includes Small Disadvantaged Business (SDB) Concerns, Women-Owned Small Business (WOSB) Concerns, Historically Black Colleges and Universities (HBCU) and Minority Institutions (MI), HUBZone Small Business (HUBZone SB) Concerns and Veteran-Owned Small Business Concerns, including Service-Disabled Veteran-Owned Small Business Concerns and Hispanic Serving Institutions and Tribal Council Universities.
Estimated Subcontracting By Reporting Category (Ref: SF294)
B. Total estimated dollar value and percent of planned expenditures for materials and supplies and certain travel expenses to be directed toward Small Disadvantaged Business (SBD) Concerns (includes HBCU/MI Concerns):
KSU hereby establishes the goal of a minimum of $10,400 of total materials and supplies and/or certain travel expenditures will be directed toward this reporting category, which includes $7,360 placed with SDB concerns and $1,520 with Historically Black Colleges and Universities and Minority Institutions and $1,520 placed with Hispanic Serving Institutions and Tribal Council Universities. These estimated allocations equate to approximately 24%, 5% and 5% respectively of all expenditures earmarked for the procurement of materials, supplies and certain travel expenses and totals 34% for this reporting category. KSU reserves the right to negotiate the modification of specific goals, as definitive procurement activities are pursued. The Project Director will work with the Kansas State University Purchasing Office and State of Kansas Purchasing offices to identify and contact small businesses that may be able to provide any necessary services.
C. Total estimated dollar value and percent of planned expenditures for materials and supplies and certain travel expenses to be directed toward Women-Owned Small Business Concerns:
KSU hereby establishes the goal of a minimum of $12,480 of total materials and supplies and/or certain travel expenditures will be directed toward this reporting category. This estimated allocation equates to approximately 41% of all expenditures earmarked for the procurement of materials, supplies and certain travel expenses. KSU reserves the right to negotiate the modification of specific goals, as definitive procurement activities are pursued. The Project Director will work with the Kansas State University Purchasing Office and State of Kansas Purchasing offices to identify and contact small businesses that may be able to provide any necessary services.
D. Total estimated dollar value and percent of planned expenditures for materials and supplies and certain travel expenses will be directed toward Historically Black Colleges and Universities (HBCU) and Minority Institutions:
KSU hereby establishes the goal of a minimum of $1,520 of total materials and supplies and/or certain travel expenditures will be directed toward this reporting category. This estimated allocation equates to approximately 5% of all expenditures earmarked for the procurement of materials, supplies and certain travel expenses. KSU reserves the right to negotiate the modification of specific goals, as definitive procurement activities are pursued. The Project Director will work with the Kansas State University Purchasing Office and State of Kansas Purchasing offices to identify and contact small businesses that may be able to provide any necessary services.
E. Total estimated dollar value and percent of planned expenditures for materials and supplies and certain travel expenses will be directed toward HUBZone Small Business (HUBZone SB) Concerns:
KSU hereby establishes the goal of a minimum of $1,520 of total materials and supplies and/or certain travel expenditures will be directed toward this reporting category. This estimated allocation equates to approximately 5% of all expenditures earmarked for the procurement of materials, supplies and certain travel expenses. KSU reserves the right to negotiate the modification of specific goals, as definitive procurement activities are pursued. The Project Director will work with the Kansas State University Purchasing Office and State of Kansas Purchasing offices to identify and contact small businesses that may be able to provide any necessary services.
F. Total estimated dollar value and percent of planned expenditures for materials and supplies and certain travel expenses will be directed toward Veteran-Owned including Service-Disabled Veteran-Owned Small Businesses:
KSU hereby establishes the goal of a minimum of $4,000 of total materials and supplies and/or certain travel expenditures will be directed toward Veteran-Owned and $2,000 to Service-Disabled Veteran-Owned small businesses, as described above. This estimated allocation equates to approximately 13% and 7%, respectfully, of all expenditures earmarked for the procurement of materials, supplies and certain travel expenses. KSU reserves the right to negotiate the modification of specific goals, as definitive procurement activities are pursued. The Project Director will work with the Kansas State University Purchasing Office and State of Kansas Purchasing offices to identify and contact small businesses that may be able to provide any necessary services.
G. Total estimated dollar value and percent of planned expenditures for materials and supplies and certain travel expenses will be directed toward Service-Disabled Veteran-Owned Small Businesses:
KSU hereby establishes the goal of a minimum of $2,000 of total materials and supplies and/or certain travel expenditures will be directed toward Service-Disabled Veteran-Owned small businesses, as described above. This estimated allocation equates to approximately 7% of all expenditures earmarked for the procurement of materials, supplies and certain travel expenses. KSU reserves the right to negotiate the modification of specific goals, as definitive procurement activities are pursued. The Project Director will work with the Kansas State University Purchasing Office and State of Kansas Purchasing offices to identify and contact small businesses that may be able to provide any necessary services.
H. Total estimated dollar value and percent of planned expenditures for materials and supplies and certain travel expenses will be directed toward Hispanic Serving Institutions and Tribal Council Universities:
KSU hereby establishes the goal of a minimum of $1,520 of total materials and supplies and/or certain travel expenditures will be directed toward Hispanic Serving Institutions and Tribal Council Universities, as described above. This estimated allocation equates to approximately 5% of all expenditures earmarked for the procurement of materials, supplies and certain travel expenses. KSU reserves the right to negotiate the modification of specific goals, as definitive procurement activities are pursued. The Project Director will work with the Kansas State University Purchasing Office and State of Kansas Purchasing offices to identify and contact small businesses that may be able to provide any necessary services.
I. Total estimated dollar value and percent planned subcontracting with large business concerns:
KSU’s goal, first and foremost, is to conduct the acquisition of subcontractual services contemplated under this contract in a manner consistent with the underlying principles that govern the Small Business Subcontracting Program. KSU will use every effort to meet and exceed the goals of this program. As mentioned in Section 2. A. and Section 2. J., KSU has proposed issuing subcontracts to American Systems Corporation, a large business concern and to The University of Mississippi, a State Non-Profit Institution of Higher Education. The anticipated amount of these subcontracts totals $1,546,813 to American Systems Corporation and $1, 174,743 to The University of Mississippi, which totals 100% of all direct subcontracting activity.
J. Provide a description of all the products and/or services to be subcontracted under this contract, and indicate the types of businesses supplying them; (i.e., SMALL BUSINESS (SB), SMALL DISADVANTAGED BUSINESS (SDB), WOMEN-OWNED SMALL BUSINESS (WOSB), LARGE BUSINESS (LARGE), etc.)
It is proposed that American Systems Corporation (ASC) and The University of Mississippi will receive contracts for subject matter expertise as detailed below under the proposed RFQ. Furthermore, it is anticipated that other line item procurement opportunities with small business concerns are anticipated, but not specifically identified.
American Systems Corporation will provide Subject Matter Expertise in urban operations, small unit capabilities and tactics, command and control (C2). Course of Action (CoA) development and assessment, and decision-making required to develop an automated Adversarial Reasoning Module (ARM) and will serve as the team’s “Northern Virginia Office” to facilitate team coordination, planning and task execution as appropriate in support of the DARPA customer located in Arlington, Virginia.
The University of Mississippi will commit expert key personnel located in its Hearin Center for Enterprise Science to conduct Bayesian Modeling, experimental design and designing optimization algorithms for simulations of urban combat situations.
K. Acquisition of certain line items expenditures such as materials and supplies, and certain travel expense will be directed toward small business concerns, as detailed above. Based on the current proposed budget, it is anticipated that of the potential budget of $152,000 for these line items, a goal of at least $30,400 will be directed toward small business concerns, which consists of 20% of the total budget for these line items.
L. Goals are consistent with purchasing guidelines of the State of Kansas and the Federal Government. Specific dollar amounts to be directed to small business concerns are estimated for the non-subcontracting expenditures.
M. Indirect costs have been included in the dollar and percentage subcontracting goals stated above.
3. Program Administrator
Name, title, position within the corporate structure, and duties and responsibilities of the employee who will administer the contractor's subcontracting program.
University-wide Contact: Bill Sesler
Director of Purchasing
Address: Kansas State University
26 Anderson Hall
Manhattan, KS 66506
Telephone: 785.532.6214
Fax: 785.532.5577
e-mail: wsesler@ksu.edu
Project Coordinator: William H. Hsu, Ph.D.
Project Coordinator
Address: Kansas State University
Department of Computing and Information Sciences
234 Nichols Hall
Manhattan, KS 66506
Telephone: 785.532.6350
Fax: 785.532.7353
e-mail: bhsu@cis.ksu.edu
Duties: Has general overall responsibility for the contractor's subcontracting program, i.e., developing, preparing, and executing subcontracting plans and monitoring performance relative to the requirements of this particular plan. These duties include, but are not limited to, the following activities:
A. Developing and promoting company-wide policy initiatives that demonstrate the company's support for awarding contracts and subcontracts to small business, small disadvantaged business, women-owned small business concerns, and Minority Serving Institutions; and assure that small business, small disadvantaged business, women-owned small business and Minority Serving Institution concerns are included on the services they are capable of providing;
B. Developing and maintaining bidder's lists of small business, small disadvantaged business, women-owned small business and Minority Serving Institution concerns from all possible sources;
C. Ensuring periodic rotation of potential subcontractors on bidder's lists;
D. Ensuring that procurement "packages" are designed to permit the maximum possible participation of small business, small disadvantaged business, women-owned small business and Minority Serving Institution concerns; within State Purchasing laws and regulations;
E. Make arrangements for the utilization of various sources for the identification of small business, small disadvantaged business and women-owned small business concerns such as the SBA's Procurement Automated Source System (PASS), the National Minority Purchasing Council Vendor Information Service, the Office of Minority Business Data Center in the Department of Commerce, National Association of Women Business Owner Vendor Information Service, and the facilities of local small business, minority and women associations, and contact with Federal agencies' Small Business Program Managers;
F. Overseeing the establishment and maintenance of contract and subcontract award records;
G. Attending or arranging for the attendance of company counselors at Small Business Opportunity Workshops, Minority and Women Business Enterprise Seminars, Trade Fairs, Procurements Conferences, etc;
H. Ensure small business, small disadvantaged business, veteran owned, service-disabled veteran owned, hubzone, and women-owned small business concerns are made aware of subcontracting opportunities and how to prepare responsive bids to the company;
I. Conducting or arranging for the conduct of training for purchasing personnel regarding the intent, requirements and impact of the various applicable Public Law’s referenced in Section 2. Goals, on KSU’s purchasing procedures, especially in regards to funds expended under this contract;
J. Monitoring the company's performance and making any adjustments necessary to achieve the subcontract plan goals;
K. Preparing, and submitting timely, required subcontract reports;
L. Coordinating the company's activities during the conduct of compliance reviews by Federal agencies;
M. Reviewing solicitations to remove statements, clauses, etc., which may tend to restrict or prohibit small business, small disadvantaged business, women-owned small business concerns participation, where possible.
N. Ensuring that the bid proposal review board documents its reasons for not selecting low bids submitted by small business, small disadvantaged business, and women-owned small business concerns.
O. Ensuring the establishment and maintenance of records of solicitations and subcontract award activity.
P. Ensuring that historically Black colleges and universities and minority institutions shall be afforded maximum practicable opportunity (if applicable).
Q. Other duties
4. Equitable Opportunity
The contractor agrees to ensure that small business, small disadvantaged business, women-owned small business and Minority Serving Institution concerns will have an equitable opportunity to compete for subcontracts. The various efforts include, but are not limited to, the following activities:
A. Outreach efforts to obtain sources:
(i) Contacting small, small disadvantaged (minority), and women-owned small business trade associations (identify specific small, small disadvantaged business, and women-owned small business trade associations).
(ii) Contacting small business development organizations (identify specific small business development organizations).
(iii) Attending small, small disadvantaged (minority), and women-owned small business procurement conferences and trade fairs (to the extent known, identify specific procurement conferences and trade fairs and dates).
(iv) Potential sources will be requested from SBA's PASS and ProNet system and SubNet, where funds are available to do so.
(v) Utilizing newspaper and magazine ads to encourage new sources.
B. Internal efforts to guide and encourage purchasing personnel:
(i) Presenting workshops, seminars, and training programs;
(ii) Establishing, maintaining, and using small business, small disadvantaged business, women-owned small business and Minority Serving Institution source lists, guides, and other data for soliciting subcontracts; and
(iii) Monitoring activities to evaluate compliance with the subcontracting plan.
C. Additional efforts:
5. Flow-Down clause
The contractor agrees to include the provisions under FAR 52.219-8, "Utilization of Small Business Concerns, Small Disadvantaged Business Concerns, and Women-Owned Small Business Concerns" in all subcontracts that offer further subcontracting opportunities. All subcontractors, except small business concerns, that receive subcontracts in excess of $500,000 ($1,000,000 for construction) must adopt and comply with a plan similar to the plan required by FAR 52.219-9, "Small Business, Small Disadvantaged Business, Women-Owned Small Business Subcontracting Plan." (FAR 19.704 (a)(4)).
Such plans will be reviewed by comparing them with the provisions of Public Law 95-507, and all other related subject matter statutes, and assuring that all minimum requirements of an acceptable subcontracting plan have been satisfied. The acceptability of percentage goals shall be determined on a case-by-case basis depending on the supplies/services involved, the availability of potential small, small disadvantaged, and women-owned small business subcontractors, and prior experience. Once approved and implemented, plans will be monitored through the submission of periodic reports, and/or, as time and availability of funds permit, periodic visits to subcontractors facilities to review applicable records and subcontracting program progress.
6. Reporting and Cooperation
The contractor gives assurance of (1) cooperation in any studies or surveys that may be required by the contracting agency or the Small Business Administration; (2) submission of periodic reports such as utilization reports, which show compliance with the subcontracting plan; (3) submission of Standard Form (SF) 294, "Subcontracting Report for Individual Contracts," and SF-295, "Summary Subcontract Report," in accordance with the instructions on the forms; and (4) ensuring that large business subcontractors with subcontracting plans agree to submit Standard Forms 294 and 295.
Reporting Period Report Due Date
Oct 1 - March 31 SF-294 04/30
Apr 1 - Sep 30 SF-294 10/30
Oct 1 - Sep 30 SF-295 10/30
7. Record keeping
The following is a recitation of the types of records the contractor will maintain to demonstrate the procedures adopted to comply with the requirements and goals in the subcontracting plan. These records will include, but not be limited to, the following:
A. If the prime contractor is not using PASS as its source for small business, small disadvantaged business, and women-owned small business concerns, list the names of guides and other data identifying such vendors;
The Purchasing Office at Kansas State University, with assistance from the State of Kansas, maintains a listing of eligible businesses. This list will be used as appropriate.
B. Organizations contacted in an attempt to locate small business, small disadvantaged business, women-owned small business sources;
C. On a contract-by-contract basis, records on all subcontract solicitations over $100,000 which indicate for each solicitation (1) whether small business concerns were solicited, and if not, why not; (2) whether small disadvantaged business concerns were solicited, and if not, why not; (3) whether women-owned small businesses were solicited, and if not, why not; and (4) reason for failure of solicited small business, small disadvantaged business, or women-owned small business concerns to receive the subcontract award;
D. Records to support other outreach efforts, e.g., contacts with minority, small business, women-owned small business trade associations, attendance at small business, minority, women-owned small business procurement conferences and trade fairs;
E. Records to support internal guidance and encouragement, provided to buyers through (1) workshops, seminars, training programs, incentive awards; and (2) monitoring of activities to evaluate compliance; and
F. On a contract-by-contract basis, records to support subcontract award data including the name, address and business size of each subcontractor. (This item is not required for company or division-wide commercial products plans.)
G. Additional records:
This subcontracting plan was submitted by:
Signature:
Typed Name: Paul Lowe
Title: Asst Vice Provost/Director
Pre-Award Services
Kansas State University
2 Fairchild Hall
Manhattan, KS 66506
Date Prepared: April 22, 2004
Phone No.: (785) 532-6804
-----------------------
[1] It may prove useful to draw a distinction between Green elements that may choose to play an active role that directly affects Blue and Red actions, perceptions, and outcomes from White elements – non-combatants and others – that do not wish to play an active role but may indirectly affect Blue and Red actions, perceptions, and outcomes (or that may be used against their wishes by Red against Blue).
[2] Real-world operations will see a divergence between perceptions and reality – there will even be differences between Blue’s perception of itself and the actual Blue state.
[3] Locally stable concepts are those that do not drift over time when restricted to or conditioned upon specific influencing context variables. This is related to the idea of context-specific independence in graphical models.
[4] ASC employs a number of senior retired flag officers from the various services and also has consulting agreements allowing ready flag officer access beyond these ASC employees.
................
................
In order to avoid copyright disputes, this page is only a partial summary.
To fulfill the demand for quickly locating and searching documents.
It is intelligent file search solution for home and business.
Related download
- v financial management kansas department of commerce
- kansas state university
- indiangrass sorghastrum nutans plant guide
- 1 little rock district u s army corps of engineers
- annual report heading 1 archives
- kansas all hazards behavioral health
- real time adversarial reasoning under uncertainty
- emergency management and homeland
- july 11 2007
- kdwpt s joe kramer recognized for wetland
Related searches
- real time options quotes free
- streaming real time stock ticker
- real time stock quotes
- real time quotes stock
- real time stock market ticker
- real time stock ticker free
- real time stock ticker
- real time streaming stock quotes
- yahoo finance real time quotes
- free real time stock charts
- free real time stock quotes streaming
- real time stock charts free