Towards a Model of Social Coherence Towards a Model of Social Coherence In Multi-Agent Organizations In Multi-Agent Organizations Erick Mart Erick Mart í í nez nez Ivan Kwiatkowski Ivan Kwiatkowski Philippe Pasquier Philippe Pasquier {emartinez, pasquier}@sfu.ca {emartinez, pasquier}@sfu.ca
21
Embed
Towards a Model of Social Coherence in Multi-Agent Organizations
We propose a social coherence-based model and simulation framework to study the dynamics of multi-agent organizations. This model rests on the notion of social commitment to represent all the agents’ explicit inter-dependencies including roles and organizational structures. A local coherence-based approach is used that, along with a sanction policy, ensures social control in the system and the emergence of social coherence. We illustrate the model and the simulator with a simple experiment comparing two sanction policies.
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Towards a Model of Social CoherenceTowards a Model of Social CoherenceIn Multi-Agent OrganizationsIn Multi-Agent Organizations
● Operational model where agent behaviour is driven by Operational model where agent behaviour is driven by tractable tractable coherence calculuscoherence calculus
● Local cLocal coherenceoherence-driven agent behaviour drives the dynamics of -driven agent behaviour drives the dynamics of multi-agent organizations; from where multi-agent organizations; from where social coherencesocial coherence emerges emerges
● Local Local coherence calculuscoherence calculus of agents incorporates of agents incorporates sanction policiessanction policies
● ImplementationImplementation
● Java-based Java-based simulation frameworksimulation framework for studying the dynamics of for studying the dynamics of social systemssocial systems
● ExperimentsExperiments
● We illustrate our model by running some We illustrate our model by running some preliminary experimentspreliminary experiments, , and contrasting two different and contrasting two different sanction policiessanction policies
Relationships inspired by TRelationships inspired by TÆÆMS' taxonomy MS' taxonomy [H[Höörling et al., 1999]rling et al., 1999]
breakOvenexogx disables cookPizza x
becomeHungry exog x enables orderPizza x makeOvenDirtyexog x hinders cookPizza x cleanKitchenVent x facilitates cookPizza x
orderPizza x enables cookPizza x cookPizza x enables deliverPizza x deliverPizza x enables payOrder xcleanOvenx disables cookPizza x repairOvenx disables cookPizza x
● Oriented responsibilities Oriented responsibilities contracted by contracted by debtordebtor towards towards creditorcreditor
● Dynamics formalized as Dynamics formalized as a finite state machine a finite state machine (FSM) (FSM) [Pasquier et al., 2006][Pasquier et al., 2006]
● Commitments can be Commitments can be manipulated (state / manipulated (state / transitions)transitions)
● Sanction-based: Sanction-based: positivepositive & & negativenegative incentives, decided incentives, decided a prioria priori, static, , static, centralizedcentralized enforcement, applied at the enforcement, applied at the time of violation time of violation
● SanctionsSanctions are embedded into the life-cycle of are embedded into the life-cycle of social social commitmentscommitments, e.g.,, e.g.,
● Agent's Agent's llevel of activityevel of activity: # of : # of acceptedaccepted ISCs in its ISCs in its agenda at any given timeagenda at any given time
● Agents cannot do more than one thing at the timeAgents cannot do more than one thing at the time
● ISCs can have ISCs can have weighted constraintsweighted constraints between them between them
● An agent will do An agent will do constraint optimizationconstraint optimization over the network over the network of ISCs (of ISCs (agendaagenda) it's involved in) it's involved in
● Coherence degreeCoherence degree: total weight of : total weight of satisfied constraintssatisfied constraints between ISC in agent's between ISC in agent's agendaagenda, divided by total weight of , divided by total weight of overall constraintsoverall constraints
G W ' = CoherenceDegree W ' − CoherenceDegree W − ResToChange x , T
where : ResToChange x , T ≡− σ SC T
● The expected utility for an agent to attempt to reach state The expected utility for an agent to attempt to reach state W' from state from state W (which (which only differsonly differs by the change of state by the change of state of a of a singlesingle ISC ISC x))
● For now, no probabilities. Decision making is myopic as For now, no probabilities. Decision making is myopic as agents only consider cancellation penaltiesagents only consider cancellation penalties
● Utility function can be improved by incorporating Utility function can be improved by incorporating uncertainty. E.g., considering probability of failure, uncertainty. E.g., considering probability of failure, rewards & penaltiesrewards & penalties
● In order to In order to maximizemaximize the the coherence degreecoherence degree of its agenda of its agenda (i.e., ISCs) an agent tries to do (i.e., ISCs) an agent tries to do constraint optimizationconstraint optimization
● Agent cycle:Agent cycle: 1. Calculate CoherenceDegree Agenda2. For each active ISC x do3. Calculate utility of flipping ISC x4. End For5. Return ISC x with higher utility gain if any
● Recursive local searchRecursive local search algorithm, algorithm, no backtrackingno backtracking, worst-case , worst-case complexity is polynomial:complexity is polynomial: O(mn2)
● n is the # of ICsis the # of ICs● m is the # of constraints between ICsis the # of constraints between ICs
Observation 3Observation 3. Under . Under SPolSPol11
the organization was the organization was more more efficientefficient than without any than without any sanctionssanctions (i.e., (i.e., SpolSpol
00). This is ). This is
because the sanction policy because the sanction policy acts as deterrence for easy acts as deterrence for easy cancellations.cancellations.
Observation 1Observation 1. Desirable . Desirable agent behaviour results from agent behaviour results from local local coherence coherence maximizationmaximization. Macro-level . Macro-level social coherencesocial coherence does does emerge from local emerge from local coherence maximizationcoherence maximization..
Observation 2Observation 2. The . The efficiencyefficiency of the organization of the organization degradeddegraded from nearly from nearly optimal as optimal as frequencyfrequency of of orders and agent's orders and agent's level of level of activityactivity was increased. was increased.
Future Work
● ModelModel (extensions): (extensions):● Introducing Introducing uncertainty reasoninguncertainty reasoning into the into the coherence coherence
calculuscalculus; reasoning about time and actions; reasoning about time and actions● Modelling agents with no knowledge, with partial knowledge, Modelling agents with no knowledge, with partial knowledge,
or with complete/shared knowledgeor with complete/shared knowledge● Machine learning mechanisms would allow agents to Machine learning mechanisms would allow agents to
progressively learn these probabilitiesprogressively learn these probabilities
● ExperimentsExperiments::● Impact of different organizational structures (e.g., Impact of different organizational structures (e.g.,
hierarchies, holarchies, societies, federations)hierarchies, holarchies, societies, federations)● Investigate other sanction policiesInvestigate other sanction policies