Summary Slide

1

Multi-agent systemen Working together
2
Cooperative Distributed Problem Solving (CDPS)

• How can a MAS task be divided into smaller tasks
• How can a solution be created for a MAS using
  results from individual agents
• What action coordination techniques can be used

3
Task sharing / Result sharing

• Task sharing a problem is decomposed into a
  number of sub-tasks (how to allocate the tasks to
  the agents?)
• Result sharing the sharing of gained information
  over the agents (how to communicate the
  information?)

4
Distributed tasks

• MAS perform a task collectively
• Therefore task distribution methods and resource
  sharing are important aspects
• A successful method is to breakdown the tasks
  into contracts

5
Allocation of tasks
6
Centralized allocation of tasks by a trader
7
Nr of messages processed by the trader vs nr of
agents
8
Centralized allocation of tasks by a trader (2)

• Advantage Simple to understand and implement
• Disadvantage Centralized source of knowledge
  about all agents

9
Distributed task allocationAcquaintance networks
Every agent keeps a list of skills of all other
agents
Agent A
A B C D (agents) C1 0 1
1 0 C2 0 0 1 0 C3 1 0 0
1 (skills)
10
Distributed task allocation1. Direct
11
Distributed allocation2. By delegation
12
Task allocation Contract nets

• Based on the market mechanism of making contracts
  between two parties
• There are 4 stages
• Request for bids to all agents for performing a
  task
• Proposals from the bidders
• Evaluation of proposals, selection of contractor
• Establishment of contract

13
Task allocation Contract nets
14
Contract nets

• Requirements
• Communication links between agents
• A common language for requesting, bidding,
  evaluation, etc.
• Limit date for proposal submission
• Commitment from the contractor

15
Contract net languagerequest for bids
16
Contract net languageproposal
17
Contract nets Ring structure
18
Contract nets Problem of sub-contracting
How long do we wait?
19
Other types of task allocation

• Combination of contract net and acquaintance
  network
• Emergent allocation

20
Emergent task allocationReactive
Gold mining robots bring gold to the user using
the shortest path method. There can be only one
agent per user (if conflict, react and find other
user or wait)
21
Volgende week

• 1. Artikelen
• 10 min (strafpunten vanaf 15min)
• In geval van een duo (20min)
• Powerpoint presentatie (inleveren)
• Geen verslag

• 2. Teambots
• 20 min (strafpunten vanaf 30 min)
• Powerpoint presentatie (inleveren)
• Verslag(inleveren)

22
Part IICoordination of actions
23
Four-stage model of CDPS(Cooperative Distributed
Problem Solving)

• 1. Recognition (by an agent that it needs others)
• 2. Team formation
• 3. Plan formation
• 4. Team action

24
Approaches to coordination between agents

• Partial global plans
• Joint intentions
• Mutual modeling
• Norms and social laws

25
Partial Global Planning
26
Joint intentions

• Based on commitments and conventions
• A commitment is a promise to do the task
• A convention is a set of rules of how to deal
  with commitments
• A commitment may be abandoned if another agent
  can do the same task (it becomes redundant)
• A group of agents share the same commitment
  (joint intention)
• Commitments and conventions can be coded as rules
  in a rule-based system

27
Example ARCHON (Jennings 1993)
28
Coordination by mutual modeling
What does the other agent want? Share the same
view of the environment (world model) Share the
same believes Optimize personal and MAS utility
(payoff matrix)
29
MACE

• MACE (by Les Gasser,1985) was the first
  experimental testbed for MAS
• MACE brought five important components together
• 1. Application agents (the actual agent
  application)
• 2. Standardized system agents (e.g. user
  interfaces)
• 3. Facilities available to all agents (software
  libraries)
• 4. Description database (maintains descriptions
  of other agents)
• 5. Kernels to handle communication, message
  passing, etc

30
Acquaintance information in MACE

• Class (agents are organized in structured groups)
• Name (of the agent)
• Roles (of the agent in its class)
• Skills
• Goals
• Plans

31
Coordination by norms and social laws

• Norms are expected patterns of behavior
• Social laws are imposed norms

32
Example Traffic management

• Cars are agents that behave using social laws
  (driving on the right lane, right goes first at
  cross-sections, etc)
• Traffic management is currently heavily studied

33
Planning methods

• Centralized planning
• There is a master agent who distributes the
  sub-plans over the agents
• Distributed planning
• Each agent generates its own plan and merges it
  with the other agents. After negotiation the
  agents execute their plans

34
Centralized planning
Operators
MoveFromTo(b,x,y) Precond on(b,x), clear(b),
clear(y) Postcond on(b,y), clear(x)
MoveToTable(b,x) Precond on(b,x),
clear(b) Postcond on(b,T), clear(x)
Given a goal description, a set of operators, and
an initial state description, generate a partial
plan order
35
Centralized planning (2)
Possible plans
S1MoveToTable(A,B) S2MoveToTable(C,D)
S3MoveToTable(E,F)
1. Decompose the plan into sub-plans 2. Check the
preconditions to select the sub-plans 3.
Synchronize the sub-plans 4. Allocate the
sub-plans to agents
S4 MoveFromTo(A,T,E) S5 MoveFromTo(D,T,C) S6
MoveFromTo(B,T,A) S7 MoveFromTo(F,T,D)
S1MoveToTable(A,B)



S3MoveToTable(E,F) S2MoveToTable(C,D)
S4 MoveFromTo(A,T,E) S5 MoveFromTo(D,T,C)
S6 MoveFromTo(B,T,A) S7 MoveFromTo(F,T,D)
36
Distributed planning
Each of the planning agents generates a partial
plan (sub-plan) in parallel Then merge these
plans into a global plan
Agent 1 is specialized in doing
MoveToTable(b,x) Agent 2 is specialized in doing
MoveFromTo(b,x,y)
Merge the plan of agent 1 with plan of agent 2 by
checking the preconditions
The agents need to communicate Typical there is
a-synchronous execution of the plans, therefore
the agents exchange handshake signals
37
Distributed planning (2)
Plan agent1
S1MoveToTable(A,B) S2MoveToTable(C,D) S3MoveToT
able(E,F)
Plan agent2
S4 MoveFromTo(A,T,E) S6 MoveFromTo(B,T,A) S7
MoveFromTo(D,T,C) S8 MoveFromTo(F,T,D)
Final Plan A1 start with S3 (best move found
after negotiation) S1 S2
Final Plan A2 Wait till A and E are free
S4 S6 Wait till D is free S7 S8
38
Aircraft tracking example
39
Volgende week

• 1. Artikelen
• 10 min (strafpunten vanaf 15min)
• In geval van een duo (20min)
• Powerpoint presentatie (inleveren)
• Geen verslag

• 2. Teambots
• 20 min (strafpunten vanaf 30 min)
• Powerpoint presentatie (inleveren)
• Verslag(inleveren)