Transcript Multi-Agent Systems Lecture 5 Computer Science WPI Spring

Multi-Agent Systems
University “Politehnica” of Bucarest
Spring 2010
Adina Magda Florea
http://turing.cs.pub.ro/mas_10
curs.cs.pub.ro
Lecture 6: Agent coordination
Lecture outline
1 Coordination strategies
2 BDI Logic
3 Modeling coordination by shared mental
states
4 A system with joint actions and
conventions
1 Coordination strategies

Coordination = the process by which an agent reasons about
its local actions and the (anticipated) actions of others to try to
ensure that the community acts in a coherent manner
Coordination
Collectively
motivated agents
common goals
Cooperation to
achieve common goal
Self-interested
agents
own goals
Coordination for
coherent behavior
Neutral to one another
disjunctive goals
Competitive
conflicting goals
3



Perfect coordination ???
Centralized coordination ?
Distributed coordination



Model
Protocol
Communication

Tightly coupled interactions - distributed search

Complex agents - distributed planning, task sharing,
resource sharing

Cooperative
Heterogeneous agents - interaction protocols: Contract
Net, KQML conversations, FIPA protocols

Dynamic interactions - Meta-level information exchange,
commitments and conventions

Complex interactions - organizational structures to
reduce complexity

Unpredictable interactions - social laws

Conflict of interests - interaction protocols: voting,
auctions, bargaining, market mechanisms, Contract
Neutral or
Net, coalition formation
competitive
4
2 BDI logic
LB - set of moment formula
LS - set of path-formula
Modal operators Bel, Des, Int, (Kw)
L I based on LB (formula) and LS(path)
A - set of agents
if pLS and x A then xBelp, xIntp, xDesp, xKwpL I
M = <W, T, <, | |, R, B, D, I>
5
2.1 BDI operators
B - belief accessible relation - belief accessible worlds; the
worlds the agent believes possible
D - desire (goal) accessible relation
 Each situation wt has associated a set of desire (goal) accessible worlds - realism
I - intention accessible relation
 Intentions - sets of intention-accessible worlds - these are
the worlds the agent has committed to realize.
 Corresponding to each goal-accessible world w at some
time t there must be an intention-accessible world that is a
subworld of w at time t
Ex: xDes(A Fwin)  xInt(E Fwork)  xBel(A Fwin)
6
intention accessible world
belief accessible world
s
r
s
s
r
s
s
p
s
q
s
q
p
s
q
r
s
r
s
q
r
s
r
s
q
goal accessible world
r - Alice is in Italy
s – Paris is the capital of France
p -Alice visits Paris
q - it is spring time
7
M |=t x Bel p iff (t': (t,t')B(x,t)  M |=t' p)
an agent x has a belief p in a given moment
t if and only if p is true in all belief accessible worlds of
the agent in that moment
M |=t x Des p iff (t': (t,t')D(x,t)  M |=t' p)
an agent x has a desire p in a given
moment t if and only if p is true in all goal accessible
worlds of the agent in that moment
M |=t x Int p iff (s: sI(x,t)  M |=s,t Fp)
at each moment t, I assigns a set of paths
that the agent x has selected or preferred, i.e., if the
agent has selected p as an intention, p will hold
eventually in the future
8
Belief-goal compatibility
If an agent adopts p as a goal, the agent believes that there is a path
on which p will be true (as it is an adopted desire)
xDesp  (xBel (E G p)
Goal-intention compatibility
If an agent adopts p as an intention, it should have adopted it as a
goal to be achieved
xIntp  xDesp
No infinite deferral
The agent should not procrastinate with respect to its intentions; if the
agent forms an intention, then sometimes in the future it will give
up this intention
xIntp A F(xIntp))
F - eventually
G - always
A - inevitable
E - optional
9
2.2 Commitments as change

Desires (goals) and intentions are quite similar in their
semantic structure.

The difference arises in their relationships with the other
modalities and in terms of how they may evolve over time.

An agent is treated as being committed to its intention but,
cf. no infinite deferral, it will give up these intentions
eventually - when?

Different types of agents will have different commitment
strategies.
10
Blindly committed agent
maintains its intentions until it believes it has achieved them
xInt(A Fp) A (xInt(A Fp)  xBelp)
Single-minded committed agent
maintains its intentions as long as it believes they are still options
xInt(A Fp) A (xInt(A Fp)  (xBelp  xBel(E Fp)))
Open-minded committed agent
maintains its intentions as long as these intentions are still its desires
(goals)
xInt(A Fp) A (xInt(A Fp)  (xBelp  xBel(E Fp  xDes(E Fp)))
F - eventually
G - always
A - inevitable
E - optional
11
3 Modeling coordination by
shared mental states
 Based on the view of intentional stance agents
 Example of intentional coordinated action
3.1 Collective mental states
Common knowledge




EGp aiGKaip
- shared knowledge
Every member in G knows EGp
E2Gp  EG(EGp)
Every member knows that every member knows that every …
Ek+1Gp  EG(EKGp)
k1
Common knowledge
CGp  p  EGp  E2Gp  …  EkGp  ...
Every member ai in group G knows p
12
Mutual belief




EGp aiGaiBelp - Every one in group G believes p - shared belief
E2Gp  EG(EGp)
Ek+1Gp  EG(EKGp) k1
MGp  EGp  E2Gp  …  EkGp  …
- Mutual belief
 Perfectly shared mental state but mutual beliefs (as common knowledge)
can not be guaranteed because communication between agents is not reliable
in terms of delivery and delay
Joint intentions (Levesque & Cohen)
C1) each agent in the group has a goal that p
aiG aiIntp (and cf goal-intentions compatibility aiIntp  aiDesp)
C2) each agent will persist with this goal until it is mutually believed that p
has been achieved or that p cannot be achieved
aiG aiInt (A Fp)  A ( aiInt(A Fp) 
(MG(Achieve p)  MG(Achieve p)))
C3) conditions (C1) and (C2) are mutually believed
MG(C1)  MG(C2)
13
Commitments
 Blindly committed agent
 Single-minded committed agent
 Open-minded committed agent
Joint commitments
Agents in the group:
the state of joint commitment is distributed

have a joint goal
the group becomes jointly committed

agree they wish to cooperate
to achieve the joint goal
Joint intentions can be seen as a joint commitment to a joint goal while in
a certain shared mental state
14
Conventions
An agent should honor its commitments provided the circumstances
do not change.
Conventions = describe circumstances under which an agent should
reconsider its commitments
An agent may have several conventions but each commitment is
tracked using one convention

Commitments provide a degree of predictability so that the agents
can take future activity of other agents in consideration when dealing
with inter-agent dependencies  the necessary structure for
predictable interactions

Conventions constrain the conditions under which commitments
should be reassessed and specify the associated actions that
should be undertaken: retain, rectify or abandon the commitment 
the necessary degree of mutual support
15
Specifying conventions
Reasons for re-assessing the commitment

commitment satisfied

commitment unattainable

motivation for commitment no longer present
Actions
R1: if commitment satisfied or
commitment unattainable or
motivation for commitment no longer present
then drop commitment
 But such conventions are asocial constructs; they do not specify
how the agent should behave towards the other agents if:
– it has a goal that is inter-dependent
– it has a joint commitment to a joint goal
16
Social Conventions
Invoke when:
 commitment dropped
 commitment satisfied
 motivation for local commitment no longer present
R1: if
local commitment to joint goal G satisfied or
local commitment to joint goal G dropped or
motivation for local commitment to joint goal G
no longer present
then inform all members jointly committed to the joint goal
17
4. A System with joint action and
conventions
GRATE System (Generic Rules and Agent model
Testbed Environment, Jennings, 1994)
ARCHON - electricity distribution management
- cement factory control
Electricity distribution management of the traffic
network
 distinguish between disturbances and pre-planned
maintenance operations
 identify the type (transient or permanent), origin and extend of
faults when they occur
 determine how to restore the network after a fault
18
3 agents
AAA - the Alarm Analysis Agent
different levels
 perform diagnosis to
BAI - the Blackout Area Identifier - identifies environment info
CSI - Control System Interface 
detects the disturbance initially and then
monitors the network evolving state
19
GRATE Agent Architecture
Inter-agent
communication
CONTROL
DATA
Cooperation &
Control Layer
Communication Manager
Acquaintance
Models
COOPERATION
MODULE
Self
Model
SITUATION
ASSESMENT
MODULE
Information
store
CONTROL MODULE
Domain Level
System
Task1
Task2
Task3
20
Agent behavior
1. Select goal and develop plan to achieve goal
2. Determine if plan can be executed individually or cooperatively
(a) joint action is needed (joint goal) or
(b) action solved entirely locally
3. if (a) then the agent becomes the organiser
3.1. Establish joint action - the organiser carries on the distributed
planning protocol (next slide)
3.2. Perform individual actions in joint action
3.3. Monitor joint action
4. if (b) then perform individual actions
5. if request for joint action then the agent becomes team-member
5.1. Participate in the planning protocol to establish joint action
5.2. Perform individual actions in joint actions
(3.2 and 5.2 adequately sequenced)
21
Establish joint action
GRATE Distributed Planning Protocol
PHASE 1
1. Organiser detects need for joint action to achieve goal G and determines that
plan P is the best means of attaining it - SAM
2. Organiser contacts all acquaintances capable of contributing to P to
determine if they will participate in the joint action - CM
3. Let L  set of willing acquaintances
PHASE 2
4. for all actions B in P do
- select agent AL to carry out action B
- calculate time tB for B to be performed
based on temporal orderings of P
- send (B, tB) proposal to A
- receive reply from A
- if not rejected then
- if time proposal modified
then update remaining actions by t
- eliminate B from P
5. if B is not empty
then repeat step 4
Agent A
1. Evaluate proposal (B, tB) against
existing commitments
2. if no conflicts then
create commitment CB to (B, tB)
3. if conflicts ((B, tB), C) and
priority(B) > priority(C)
then create CB and reschedule C
4. if conflicts ((B, tB), C) and
priority(B) < priority(C)
then
if freetime (tB+ t)
then note CB and return (tB+ t)
else return reject
22
Joint intention - Phase 1 for agent AAA
Name: Diagnose-fault
Motivation: Disturbance-detection-message
Plan: { S1: Identify_blackout_area, S2: Hypothesis_generation,
S3: Monitor_disturbance, S4: Detailed_diagnosis, S2 < S4}
Start time:
Maximum end time:
Duration:
Priority: 20
Status: Establish group
Outcome: Validated_fault_hypothesis
Participants: ((Self organiser agreed_objective)
(CSI team-member agreed_objective)
(BAI team-member agreed_objective))
Bindings: NIL
Proposed contribution:
((Self (Hypothesis_generation yes) (Detailed_diagnosis yes))
(CSI (Monitor_disturbance yes)
(BAI (Identify_blackout_area yes)))
23
Joint action - Phase 2 for agent AAA
Name: Diagnose-fault
Motivation: Disturbance-detection-message
Status: Establish plan
Start time: 19
Maximum end time: 45
Bindings: ((BAI Identify_blackout_area 19 34)
(Self Hypothesis_generation 19 30)
(CSI Monitor_disturbance 19 36)
(Self Detailed_diagnosis 36 45))
….

BAI's individual intention for producing the blackout area
Name: Achieve Identify_blackout_area
Motivation: Satisfy Joint Action Diagnose-fault
Start time: 19
Maximum end time: 34
Duration: 15
Priority: 5
Status: Pending
Outcome: Blackout_area
24
Monitor the execution of joint action
Recognize situations that change commitments and impact joint
action
R1match: if task t has finished executing and
t has produced the desired outcome of the joint action (assoc to G)
then the joint goal G is satisfied
R2match: if receive information i and
i is relevant to the triggering conditions for joint goal G and
i invalidates beliefs for wanting G
then the motivation for G is no longer present
Social conventions
R1inform: if the joint goal G is satisfied
then inform all team members of successful completion of G
R2inform: if motivation for joint goal G is no longer present
then inform other team members that G needs to be abandoned
25
References





Multiagent Systems - A Modern Approach to Distributed Artificial
Intelligence, G. Weiss (Ed.), The MIT Press, 2001, Ch.2.3, 8.5-8.7
V.R. Lesser. A retrospective view of FA/C distributed problem solving.
IEEE Trans. On Systems, Man, and Cybernetics, 21(6), Nov/Dec
1991, p.1347-1362.
N.R. Jennings. Coordination techniques for distributed artificial
intelligence. In Foundations of Distributed Artificial Intelligence, G.
O'hara, N.R; Jennings (Eds.), John Wiley&Sons, 1996.
N.R. Jennings. Controlling cooperative problem solving in industrial
multi-agent systems using joint intentions. Artificial Intelligence 72(2),
1995.
E.H. Durfee. Scaling up agent coordination strategies. IEEE
Computer, 34(7), July 2001, p.39-46.
26