Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute...

20
Collectively Cognitive Collectively Cognitive Agents Agents in Cooperative Teams in Cooperative Teams Jacek Brzeziński, Piotr Dunin- Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Institute of Informatics, Warsaw University

Transcript of Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute...

Page 1: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Collectively Cognitive Agents Collectively Cognitive Agents in Cooperative Teamsin Cooperative Teams

Jacek Brzeziński, Piotr Dunin-KępliczInstitute of Computer Science, Polish Academy of Sciences

Barbara Dunin-KępliczInstitute of Computer Science, Polish Academy of Sciences

Institute of Informatics, Warsaw University

Page 2: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

GoalsGoals

Simulator– Collective Commitment Theory (B. Dunin-Kęplicz, Verbrugge)

– Theory of Trust (Castelfranchi, Falcone)

Research of some BDI systems propertieswhen different definitions of collective commitments are used:– are agents willing to cooperate?– does the system evolve? – quality of teamwork?– how much information is needed for agents to

cooperate productively?– how to calculate the degree of trust?

Page 3: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

OverviewOverview

Collective Commitments theorySimulator:– agents– interaction scenario– different CC models w.r.t. trust

Test resultsFuture work

Page 4: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Collective commitmentsCollective commitments

Tuning machine allows to select collective commitment type w.r.t.:– group awareness about social plan– group awareness about distribution of

bilateral commitments– existence of collective intention

Different models of organizations

Page 5: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Examples of collective commitmentExamples of collective commitment

Robust collective commitment

Weak collective commitment

Distributed commitment

))),,((BEL-C(

)),,(C()),((BEL-C

),()(INT-C)(COMM-R

G,

,G

G,

jiCOMM

jiOMMPconstitute

Pconstitute

GjiP

GjiP

PG

)),,((BEL-C

)),,(C(),()(INT-C)(COMM-W

,G

,G,

jiCOMM

jiOMMPconstitute

GjiP

GjiPPG

)),,(C(),()(COMM-D ,, jiOMMPconstitute GjiPPG

Page 6: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Roles of agentsRoles of agents

Managers– generate tasks (sets of actions) to be

performed by Workers for a specific salary– do not execute actions– delegate actions (or sets of actions) to other

agents (Workers)

Workers– perform actions– cannot delegate actions

Page 7: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Properties of agentsProperties of agents

Beliefs– believes(A, f) - agent A believes f

Intentions– intends(A, a) - agent A intends to perform action a

Abilities– able(A, a, n) - agent A is able to perform action a

(n - probability of success)

Trust– trusts(A, B, a, n) - agent A trusts in agent B to do a in degree of n

– trust values base on direct experience.

– Workers & Manager trust in Workers to execute specific actions.

– Workers trust in Manager to do proper team selection.

Page 8: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Four levels of CPSFour levels of CPSManager has a task (set of actions) to perform:1. Potential recognition

Manager recognizes potential for cooperative action in order to complete the task

2. Team formationManager attempts to establish group of agents that can collectively fulfil the goal (collective intention)

3. Plan generationSocial plan of achieving the goal is built and the collective commitment is formed

4. Team actionAgents involved do their tasks and eventually achieve the main goal

Page 9: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Potential recognitionPotential recognition

M sends CFP to all the Workers in the system (content = (O - set of actions))

Workers respond with their bids M chooses the best group G that can

perform the task collectively w.r.t.:– trust in Workers in the group– prices proposed by Workers – risk factor describing preferences of M

Page 10: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Team formation & plan generation 1Team formation & plan generation 1

M broadcasts relevant information to all Workers in G:– an offer to form a group– an offer to participate in task– an allocation of actions– division of task O into subtasks– an offer to perform action(s) for price x

Page 11: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Team formation & plan generation 2Team formation & plan generation 2

Agents decide whether to join the group Task result estimation function takes into account following factors:• trust in the manager M• trust in other members of the group• ability to perform delegated actions

If all Workers agree, then collective commitment is constructed.

Page 12: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Team actionTeam action

Workers perform actions and tell M about results

If all actions have succeeded, the task also succeeds and all Workers collect their payments

Results of actions are propagated to group members, depending on team awareness level

Agents modify their degrees of trust in other group members (basing on direct experience)

Page 13: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Robust collective commitmentRobust collective commitment

Information sent to Workers– an offer to create a group G with intention to realize

task O– an allocation of sets of actions– an offer to perform a set of actions for a price of x

Task result estimation function

Result propagation– All the information about successes and failures

},\{

),,DoT(*),able(*

*)"",,DoT(

MAGI SaSa IA

aIAaA

ionteamselectMA

Page 14: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Weak collective commitmentWeak collective commitment

Information sent to Workers– an offer to create a group G with intention to realize

task O– a division of task O into subtasks S– an offer to perform a set of actions for a price of x

Task result estimation function

Result propagation– only failure information: agent, failed action, subset of O

delegated to the agent

AA SOa MAGISa

aIAG

aA

ionteamselectMA

\ },\{

)),,DoT(2||

1(*),able(*

*)"",,DoT(

Page 15: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Distributed commitmentDistributed commitment

Information sent to Workers– an offer to perform a set of actions for a price

of x Task result estimation function

Result propagation– only the team action result

ASa

aAable

ionteamselectMA

),(*

*)"",,DoT(

Page 16: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

TestsTests

Configuration• 16 Workers, 1 Manager• 100,000 tasks• 4 actions in every task• abilities set randomly between 0.5 and 1• starting trust value = 1

Parameters• number of successful/failed contracts• efficiency• difference between trust values and real abilities

Page 17: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Test ResultsTest Results

Strong models of CC:– efficient–moderate number of contracts– agents learning from experience– costly (comm. resources)

Weak models– unstable–many contracts, mostly failing

Page 18: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Robust Collective CommitmentRobust Collective Commitment

Page 19: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Distributed CommitmentDistributed Commitment

Page 20: Collectively Cognitive Agents in Cooperative Teams Jacek Brzeziński, Piotr Dunin-Kęplicz Institute of Computer Science, Polish Academy of Sciences Barbara.

Future workFuture work

Complex organization model– more managers– more detailed trust model– more complex commitment structure– institutional restrictions

Reconfiguration– role of trust

Dynamic commitment model evolution