Strategic Research Directions in AI: Distributed AI and Agent Systems

1
Strategic Research Directions in AIDistributed
AI and Agent Systems

• Edmund H. Durfee
• University of Michigan AI Laboratory

2
Agent Coordination and Control
Should one or more agents notice?
Should one or more agents respond?
3
Technical ChallengeConstrained Real-Time
Responsiveness

• Cannot be prepared to notice and respond to
  everything at once.
• Goal Maximize real-time responsiveness under
  operational and execution constraints.
• Example Technical Foundations temporal
  constraint networks and MDPs.
• Example Strategic Direction Constrained MDPs
• Formulation constraints Limited computational
  resources and/or time to formulate policies.
• Operationalization constraints Platform
  resources (perception, actuation, memory) limit
  size/complexity of policies.
• Execution constraints Consumable resources
  (power, fuel, bandwidth) limit duration/persistenc
  e of policies.

4
Constrained vs Unconstrained MDPs
5
Technical ChallengeBoundedly-Optimal MultiAgent
Systems

• Responsiveness responsibilities can be
  distributed across multiple agents.
• Goal Maximize system-wide real-time
  responsiveness under individual and collective
  constraints.
• Example Technical Foundations single-agent
  bounded optimality techniques and MDPs.
• Example Strategic Direction Multiagent MDPs
  co-designing for coordination protocols and
  individual decision-making
• Growing activity on MAMDPs (UMass, USC, Toronto,
  UMich,)
• Example Iterated convergence on compatible
  schedulable policies across agents.

6
Convergence Protocol Example

• Represent possible (re)actions of other agents
  as temporal transitions (ttac labels).
• Both agents may handle dangerous state D.

7
Convergence Protocol Example
Knowing which reactions other agent plans can
restrict which states this agent must worry about.
8
Convergence Protocol Example

• Revelation of choices through protocol may
  eliminate entire subspace (and hence need to
  plan/schedule actions for those states).

9
Convergence Protocol Example

• Revelation of choices through protocol may
  eliminate just a required action, by knowing
  other agent will handle the important event.

10
Technical ChallengeSocial Autonomy

• Interdependence comes at a cost.
• Uncertainty over commitment fulfillment, and even
  definition!
• Goal Automate striking an informed and flexible
  balance between risks and benefits of dependence.
• Example Technical Foundations reflective
  architectures, adjustable autonomy, abstraction,
  commitment/convention.
• Example Strategic Direction Modeling methods,
  protocols, and languages for coordination with
  strategic ignorance.
• Revealing too many details incurs overhead.
• Revealing too many details reduces local
  flexibility.
• Revealing too few details encourages
  inefficiencies.
• Revealing too few details increases risk.

11
Tradeoffs in Knowledge Revelation

• Each agent has a couple of routes from which to
  choose.

A
A
B
B
12
Tradeoffs in Knowledge Revelation

• Not revealing any information can be risky.

A
B
13
Tradeoffs in Knowledge Revelation

• Revealing specific plans could remove some kinds
  of risk, but could jeopardize success.

14
Tradeoffs in Knowledge Revelation

• Remaining vague about plans retains flexibility,
  but can reduce efficiency (less parallelism)

A
B
15
Technical ChallengeRelationship Discovery

• In larger, more emergent multiagent systems,
  relationships happen in often unexpected ways.
• Goal Discover, represent, and resolve important
  relationships cost-effectively in sparsely
  interacting contexts.
• Example Technical Foundations forwarding
  protocols, summarization, self-description
  languages, graphical games, multiagent learning,
  distributed CSPs.
• Example Strategic Direction Emerging
  aggregations/coalitions formed around a
  discovered commonality.
• Broader communication of behavior abstractions
• Narrower drill-down into specific relationships
• Modeling in a sparse graphical representation
• Resolution one-shot, or persistent (organization)

16
Hierarchical Representation
A
A to destination
B
A lower route
A upper route
A up
A down
A down
A up
17
Top-Down Coordination
18
Top-Down Coordination
19
Top-Down Coordination
20
Top-Down Coordination
21
Top-Down Coordination
22
Top-Down Coordination
23
Top-Down Coordination
24
Top-Down Coordination
25
Top-Down Coordination
26
Technical ChallengeManagement in Continuous
Operations

• In larger, more emergent multiagent systems, the
  world evolves in often expected ways.
• Goal Manage activities and interactions to adapt
  to changing circumstances and exploit ephemeral
  opportunities.
• Example Technical Foundations dynamic belief
  networks, disjunctive temporal constraint
  networks, stalling strategies, plan repair,
  conditional planning.
• Example Strategic Direction Multiagent Plan
  Management.
• Distributed disjunctive temporal constraint
  networks
• Distributed dynamic belief networks
• Exploitation of graphical models for selective
  information propagation and processing

27
Technology Gaps

• Can (somewhat) do
• MDPs, POMDPs
• Reactive behavior
• Plan for achievement
• Adjustable autonomy
• Relationship identification
• One-shot coordination

• Need to do
• Distributed, constrained MDPs
• Real-Time behavior
• Manage continuous plans
• Social autonomy
• Large-scale coalescing into relationship clusters
• Emergent organization

28
Other Directions

• Languages and Ontologies
• Deception/Trust
• Problem Decomposition and Distribution
• Team Composition and Coordination
• Resource Allocation
• Multiagent Learning

29
Agent Coordination and Control Challenges and
Solutions