An Integrated Approach to Decision Making under Uncertainty UCLA: A' Darwiche, W' Karplus , P' Kellm

1
An Integrated Approach to Decision Making under
UncertaintyUCLA A. Darwiche, W. Karplus , P.
Kellman, J. PearlUCI R. Dechter, S.
IraniUIUC D. Roth
2
Project Objectives

• Develop basic methods for helping human-decision
  makers attain their full potential in uncertain
  environments
• Integrate the developed methods into a
  decision-aiding system, to illustrate their
  utility in enhancing the decision making process

3
Player
Player
Player
Centralized Computer
Player
Player
Decision Maker
4
(No Transcript)
5
Rules of Conduct

• The goal of the game is to gain victory-points by
  conquering and controlling other territory using
  a combination of military might and diplomacy.
• Can contain from 2 to 20 players.
• 21 different types of military units broken into
  three categories land, sea, and air.
• Each player has 6 types of resources (food, fuel,
  heavy-metal, light-metal, credits, and production
  units).
• Diplomacy can be officially accomplished by
  declaring alliances, by declaring wars, and by
  declaring neutrality.

6

• Battles occur whenever two enemy countries have
  military units in the same map area.
• The outcome of each battle is determined
  probabilistically based on each unit type in the
  area and the target unit type.
• There are "to-hit" tables in the rules that give
  probabilities that a given unit-type can hit
  another unit-type.
• There are restrictions about where airplanes are
  launched from and land as well as where naval
  units can be.
• ..
• About 40-50 pages describing rules of conduct

7
Sources of Uncertainty

• Outcome of battles fought
• Uncertain/partial intelligence of opponents
• Damage being done to ports/forts/airbases when
  attacking an area
• Amount of resources left after an area is
  conquered
• Enemy/ally's reliability
• When the game will end

8
Player
Player
Player
Centralized Computer
Player
Player
Decision Maker
Decision-Aid
9
Centralized Computer
Player
Player
Player
Decision Maker
Situation Model
Interface
Causal Queries
Inference Engine
Player
Decision-Aid
10
(No Transcript)
11
(No Transcript)
12
Three Levels of Decision-Making Support

• Compute a full (conditional) plan
• Compute a single decision
• Compute answers to queries which are needed to
  make informed decision

13
Requires Integration of Techniques for..

• Representing uncertain, incomplete information
• Fusing uncertain information of different kinds
• Evaluating and ranking courses of actions
• Providing real-time incremental responses to
  user queries
• Interacting with users in cognitively grounded
  manner

14
Elements of Research Program

• Probability theory
• Knowledge representation and reasoning
• Algorithms
• Natural language processing
• Machine learning
• Cognitive Science/Psychology

15
Key Commitments

• Probability theory as the foundation for managing
  uncertain information
• Bayesian belief networks as the mechanism for
  realizing computer implementations of uncertainty
  methods
• Mix of planned theoretical developments and
  practical implementations

16
Project Tasks

• Task A Representation and Integration of
  Uncertain Information
• Task B Course-of-Action Evaluation
• Task C Advanced Inference Techniques
• Task D User Interaction
• Task E Building an Integrated Decision-Aid

17
Task A Representation and Integration of
Uncertain Information

• Objective Develop basic methods for constructing
  a situation model that integrates diverse pieces
  of information, which can be quite dynamic and
  fraught with uncertainty and incompleteness.
• Elements Probability theory, knowledge
  representation and reasoning, machine learning.

18
Task A Representation and Integration of
Uncertain Information

• Challenges Coherent and efficient extension of
  Bayesian networks to accommodate diverse types of
  information.
• Subtasks
• Temporal information
• Constraint-based information
• Incomplete information

19
Task B Course-of-Action Evaluation

• Objective Develop a causal query language to be
  used by decision makers in inquiring about the
  relative merits of alternative courses of actions
  (COAs) in light of uncertain and incomplete
  information.
• Elements Probability theory, KR, philosophy,
  psychology.

20
Task B Course-of-Action Evaluation

• Challenges Designing the query language,
  equipping it with the appropriate semantics so it
  can be used to phrase novel queries which can
  bring new insights into COAs and their embedding
  situations.
• Subtasks
• Production and sustenance
• Actual causation
• Plan metrics

21
Task C Advanced Inference Techniques

• Objective Provide a computational engine for
  processing queries of Task B efficiently.
• Elements Mostly algorithmic computer science,
  although it does intersect with cognitive
  science/psychology

22
Task C Advanced Inference Techniques

• Challenges Dynamic Bayesian networks, meaningful
  notions of approximation, on-line computations
• Subtasks
• Any-time/approximate inference
• Real-time inference
• Incremental/robust inference

23
Task D User Interaction

• Objective Enhance situation awareness through
  the manner in which information is presented to a
  decision maker---row processed information from
  Tasks A-C.
• Elements Natural language processing, cognitive
  science, learning theory, and psychology

24
Task D User Interaction

• Subtasks
• Free-style adaptive interface requesting
  in-depth information in a more sophisticated,
  less restricted style.
• Cognitively grounded interface interface design
  based on studies from cognitive science
• Cognitive illusions
• Perceptual learning

25
Task E Building an Integrated Decision-Aid

• Objective
• Develop an integrated decision-aid based on the
  methods of Tasks A-D
• Illustrate its effectiveness in aiding a
  decision-maker in the context of engaging a
  selected class of commercially available computer
  games
• Challenges Choice of computer games, development
  of an experimental methodology

26
Computer Games as Simulation Environments

• Published APIs, text-based interfaces
• Many game genres varying levels of complexity,
  different emphasis on issues
• Precise scoring models
• Accessible to researchers
• Semi-realistic domains

27
Concluding Remarks

• Decision-Aid not a decision maker, not a
  planner.
• Two elements to bringing insights into a
  situation
• Reasoning about subtle aspects of a situation
• Appropriate user interfaces to available
  information
• Emphasis on integration
• Applicability to commercial domains medical
  applications, nuclear reactor air traffic
  control

28
Team Briefings

• Pearl, Causal reasoning for decision aiding
  systems
• Darwiche, Scaling up inference in uncertainty
  models
• Dechter, Integrating probabilistic and
  deterministic information
• Irani, On line algorithms for incremental and
  robust inferences
• Karplus, Toward more effective multi-media
  interfaces for human interaction with
  computer/communication systems
• Kellman, Optimizing interfaces and training in
  rich information situations
• Roth, Robust natural language based
  human-computer interaction A learning centered
  approach