Reasoning in Uncertain Adversarial Environments in AgentMultiagent Systems - PowerPoint PPT Presentation

Title: Reasoning in Uncertain Adversarial Environments in AgentMultiagent Systems


1
Reasoning in Uncertain Adversarial Environments
in Agent/Multiagent Systems

• Praveen Paruchuri
• University of Southern California

Guidance Committee Milind Tambe, Leana
Golubchik, Gaurav S. Sukhatme, Sarit Kraus, Stacy
Marsella, Fernando Ordonez
2
Motivation The Prediction Game

• An UAV (Unmanned Aerial Vehicle)
• Flies between the 4 regions
• Can you predict the UAV-fly patterns ??
• Pattern 1
• 1, 2, 3, 4, 1, 2, 3, 4, 1, 2, 3, 4,
• Pattern 2
• 1, 4, 3, 1, 1, 4, 2, 4, 2, 3, 4, 3, (as
  generated by 4-sided dice)
• Can you predict if 100 numbers in pattern 2 are
  given ??
• Randomization decreases Predictability
• Increases Security

3
Problem Definition

• Problem Provide security for agent/agent-team
  acting in uncertain adversarial environments.
• Assumptions for Agent/agent-team
• Acting in uncertain adversarial environments
• Adversary is unobservable
• Adversarys actions/capabilities or payoffs are
  unknown or difficult to model explicitly
• Assumptions for Adversary
• Can see the agents state or belief state
• Knows the agents plan/policy
• Exploits the action predictability

4
Solution Technique

• Technique developed
• Intentional policy randomization for security
• MDP/POMDP framework to handle sequential decision
  making under environment uncertainty
• MDP ? Markov Decision Process
• POMDP ? Partially Observable MDP
• Increase Security gt Solve Multi-criteria problem
  for agents
• Maximize action unpredictability (Policy
  randomization)
• Maintain reward above threshold (Quality
  constraints)

5
Domains

• Scheduled activities at airports like security
  check, refueling etc
• Observed by terrorists
• Randomization of schedules helpful
• UAV/UAV-team patrolling humanitarian mission
• Adversary monitoring UAV schedule to disrupt
  mission
• Can disrupt food, harm refugees, shoot down UAVs
  etc
• Randomize UAV patrol policy (More domains in
  report)

6
My Contributions

• Two main contributions
• Single Agent Case
• Formulate as Non linear program Entropy based
  metric
• Convert to Linear Program called BRLP ( Binary
  search for randomization )
• Randomize single agent policies with reward gt
  threshold
• Multi Agent Case RDR ( Rolling Down
  Randomization )
• Randomized policies for decentralized POMDPs
• Threshold on team reward

7
Related work

• Randomized policies in literature
• Stochastic Games
• Rock-Paper-Scissors game, Littman 1994
• CMDPs
• Constrained MDPs, Altman 1999
• Privacy
• Act on information while maintaining privacy,
  Otterloo 2005
• Security via randomization
• Patrol units for security, Carroll 2005
• Randomized patrol sentry vehicles, Lewis 2005
• Randomized Police Patrol, Billante 2003

8
Plan for Talk

• Achieved Contribution versus Expected Contribution

Increase Security for Agent/Agent Team Acting in
Uncertain, Adversarial domains
No Adversary Model
Partial Adversary Model
No Communication
Limited Communication
Dec- POMDP based Agent Team
Dec-MDP based Agent Team
Dec-POMDP Based Agent Team
MDP based Single Agent
9
MDP based single agent case

• MDP is tuple lt S, A, P, R gt
• S Set of states
• A Set of actions
• P Transition function
• R Reward function
• Basic terms used
• x(s,a) Expected times action a is taken in
  state s
• Policy (as function of MDP flows)

10
Entropy Measure of randomness

• Randomness or information content quantified
  using Entropy ( Shannon 1948 )
• Entropy for MDP -
• Additive Entropy Add entropies of each state
• Weighted Entropy Weigh each state by it
  contribution to total flow
• where alpha_j is the initial flow of the system

11
Tradeoff Reward vs Entropy

• Non-linear Program Max entropy, Reward above
  threshold
• Objective (Entropy) is non-linear
• BRLP ( Binary Search for Randomization LP )
• Linear Program
• No entropy calculation, Entropy as function of
  flows

12
BRLP

• Given input and target reward
• Poly-time convergence to within of target
  reward.
• Monotonicity Entropy decreases or constant with
  increasing reward.
• Control through
• Input can be any high entropy policy
• One such input is the uniform policy
• Equal probability for all actions out of all
  states
• Controls the final policy structure

13
LP for Binary Search

• Policy as function of and
• Linear Program

14
BRLP in Action
Beta .5
1 - Max entropy
0 Deterministic Max Reward
Target Reward
15
Results (Averaged over 10 MDPs)

• For a given reward threshold,
• Highest entropy Expected Entropy Method 10
  avg gain over BRLP
• Fastest BRLP 7 fold average speedup over
  Expected Entropy
• These results are statistically significant
  (t-Tests performed)

16
Multi Agent Case Problem

• Maximize entropy for agent teams subject to
  reward threshold
• For agent team
• Decentralized POMDP framework used
• Agents know initial joint belief state
• No communication possible between agents
• For adversary
• Can calculate the agents belief state
• Knows the agents policy
• Exploits the action predictability
• For Dec-POMDP-
• Deterministic policy maps observation history to
  action.
• Randomized Policy maps observation history to
  action probability distribution.

17
Policy trees Deterministic vs Randomized
Deterministic Policy Tree
Randomized Policy Tree
18
RDR Rolling Down Randomization

• Input
• Best ( local or global ) deterministic policy
• Percent of reward loss
• d parameter Number of turns each agent gets
• Ex d .5 gt Number of steps 1/d 2
• Each agent gets one turn ( for 2 agent case )
• Single agent MDP problem at each step
• For agent 1s turn
• Fix policy of other agents ( Agent 2 )
• Find randomized policy
• Maximizes joint entropy ( w1 Entropy(agent1)
  w2 Entropy(agent2) )
• Maintains joint reward above threshold

19
RDR d .5
Agent 1 Maximize joint entropy Joint Reward gt 90
Max Reward
Reward 90
Agent 2 Maximize joint entropy Joint reward gt 80
80 of Max Reward
20
RDR details

• For single agent sufficient
  statistic
• Not sufficient for multi agent case
• For multiagent case with
• Deterministic policy (other agents policies
  fixed)
• Reason about current world state observation
  history of other agents
• Randomized policy
• Current world state, action and observation
  history of other agents
• Define extended state
• where
• Joint belief state is a probability distribution
  over extended states.

21
New Transition and Observation functions

• Transition function for deterministic case
• Transition function of RDR
• Observation function remains the same

22
Belief Update Rule

• Original Belief Update Rule
• New Belief Update Rule

23
Experimental Results Reward Threshold vs
Weighted Entropy ( Averaged 10 instances )
24
Future work Dec-MDPs with Bandwidth Constraints

• Agents teams can communicate
• Limited bandwidth assumed
• Bandwidth modeled as a shared resource constraint
• Typical policy randomization formulation

25
Dec-POMDPs with Bandwidth Constraints

• RDR assumes no communication
• Communication in Dec-POMDPs
• With Deterministic Policies Nair et al, 04
• Communicate observation histories
• Compress belief states and increase expected
  rewards
• With Randomized Policies
• Can communicate observation histories
• Or communicate action taken
• Optimization Problem Optimal allocation of
  bandwidth between observation histories and
  actions st Expected reward maximized.

26
Incorporating Adversary Models

• Real world situation
• Known Adversary cannot target leftmost region of
  mission
• No UAV patrolling needed there
• Prior knowledge to be incorporated while
  randomizing policy.
• Partial adversary models Standard framework
  needs to be developed.
• Optimal plan for modeled part of adversary
• Plan for the unmodeled part also.

27
Timeline
Defense
Thesis Writing
Experimental Evaluation
Incorporating Adversarial Models
Experimental Evaluation
Dec-MDP/POMDP with communication constraints
Proposal
March 06
April 06
June 06
August 06
November 06
Feb 07
March 07
28
Other Work

• Self-interest vs Team-interest
• Electric Elves Domain
• Teamwork with resource constraints
• Developed the EMTDP framework
• Maximize expected team reward while bounding
  expected resource consumption
• CRLP (Convex Combination for Randomization)
  algorithm
• Heuristic algorithm for single agent policy
  randomization
• Communication issue in Dec-MDPs
• Communication increases security

29
Summary

• Intentional randomization as main focus
• Single agent case
• BRLP algorithm introduced
• Multi agent case
• RDR algorithm introduced
• Multi-criterion problem solved that
• Maximizes entropy
• Maintains Reward gt Threshold

30
Thank You

• Any comments/questions ??

31
(No Transcript)