1. 1 Simulating Computational Societies Lloyd Kamara, Alexander Artikis, Brendan Neville, Jeremy Pitt Imperial College, London 16-17 September 2002, Universidad Rey Juan Carlos, Madrid, Spain

2. 2 Situated Electronic Society Agent A Agent B Organization X Organization Y Accountability Information Trading (Institutionalised) (Decentralised) Agent Society Human Society Contract Situatedness AccountabilitySoftware accountability Market EconomyInformation trading No universal truthDecentralisation Rule of Law Institutionalised Int’n

3. 3 Contents  Viewing multi-agent systems (MAS) from three perspectives:  agent architecture  social decision-making of an individual agent  external observation of MAS as a whole owner states of society society formation & society dynamics situated society legal consequences & legal transfer communication protocols & socio-cognitive relationships social roles communication protocols & socio-cognitive relationships social roles states of society society formation & society dynamics communication protocols & socio-cognitive relationships social roles communication protocols & socio-cognitive relationships social roles

4. 4 Interaction Perspectives  Agent architecture  Abstract/instance for interaction with environment  provides how, but not why or when mechanism for reasoning and communication  Customisable and configurable  Socio-cognitive reasoning and relationships  for ‘supra-functional requirements’ of interactions  Social control appropriate to rights management  (e.g.) trust and reputation inform/direct behaviour  External observation  Institutional view of interactions  Characterisation of dynamic normative positions  eschews internals, admits verification (on-line & off-line)

5. 5 An Information Trading Society  The Abstract Producer/Consumer Scenario  Producers provide one resource to Consumers in exchange for another  scope for competition, co-operation, subterfuge, prohibition, punishment, …  Roles dictate capabilities of participants  Producers initiate Contract Net Protocols  Consumers initiate Auctions Instantiated by the Oil Exploration Scenario

6. 6 Scenario Description network cloud 1..m cartographers 1..n explorers auction house regulator certification authority bank 1..k rights managers

7. 7 The Simulation Model  Individual processes to represent cartographer and explorer agents  Allowance for heterogeneous agents (in terms of internal architecture)  Communications mechanism supporting multiple, simultaneous interactions  Control & configuration options for each agent.

8. 8 Agent Architecture A generic, configurable engine for experiments

9. 9 Socio-Cognitive Software Agents  A simulation of a open agent system E- commerce application  Abstract producers & consumers interacting through a contract net protocol variant.  A socio-cognitive model,  Based on the model by Castelfranchi and Falcone  Implemented socio-cognitive agents  Simulated small groups of these agents

10. 10 Socio-cognitive Beliefs  Trust  Is the degree to which Agent A trusts Agent B to complete a task, it is represented by a subjective probability.  Trust is formed as a function of reputation and direct experience.  Direct Experience (subjective probability)  the belief one agent has about the trustworthiness of another, based purely on first-hand interactions.  Reputation (subjective probability)  the belief one agent has about the reputation of another, based upon the testimonies of its peers.

11. 11 Agent Configuration  Parameterisation  Some parameters are needed as input  into the belief update functions  and Economic mechanisms.  others help to define the behaviour of the agent,  E.g. their ability is a probability of successful completion of an attempted action.  Characterisation  Co-operators only choose cooperation strategies  Defectors choose based upon maximizing the expected pay-off  Reciprocators have probabilistic intentions; the probability that they will co-operate (their trustworthiness) is matched to their evaluation of a trading partner's trustworthiness.

12. 12 Experiment  We now discuss an example simulation for trust update based on direct experience.  The experiment consists of twelve agents, split evenly between producers (explorers) and consumers (cartographers).  All of the agents are of reciprocator character type  They differ only in their ability to complete the contracted tasks.

13. 13 Experimental Results  Plots the level of trust directed towards groups of members  The featured DoT is simply the mean of the opinions of all the members in the society. Based on direct experience.

14. 14 Experimental Results  shows the performance measure, the amount of assets accrued or lost.  From this it can be seen that a society of reciprocators is a meritorious one.

15. 15 Executing the Specifications of Computational Societies  Computational societies are viewed from an external/global perspective  Computational societies as instances of normative systems:  Need to represent what the members are permitted to do, obliged to and other more complex normative relations between agents.  Need to represent the institutionalised powers of the members; a standard feature of all norm-governed organisations/societies.  Need to formally represent and reason about the actions of the members

16. 16 Theoretical Framework  Social constraints  Three levels of specification:  What kind of actions ‘count as’ valid, well-formed actions. Distinguishing between valid and invalid actions enables the separation of meaningful from meaningless activities.  What kind of actions are permitted. Determining the permitted, prohibited, obligatory actions enables the classification of the agent behaviour as legal or illegal, ethical or unethical, social or anti-social, etc.  What are the sanctions and enforcement policies that deal with illegal, unethical, anti-social behaviour.

17. 17 Computational Framework SOCIAL CONSTRAINTS: initiates(a1, pow(agentB,a3), Time). terminates(a1, pow(agentA,a1), Time). holdsAt(permitted(Agent, Act), Time):- holdsAt(pow(Agent, Act),Time). … NARRATIVE: happens(a1, t). happens(a2, t). … CURRENT SOCIAL STATE: -holdsAt(pow(agentA, a1), t+1). holdsAt(pow(agentB, a3), t+1). holdsAt(permitted(agentB,a3), t+1). …  Given (a temporal ordering of) all externally observable actions, it will provide a (graphical) representation of the institutional powers, permissions and sanctions of the members of the simulated societies (at each point in time) for the benefit of the designers and the members.

18. 18 Summary & Discussion  Simulating norm-governed computational societies: architectural, socio-cognitive, external perspective  What this offers:  Evaluation of architectural design, trust update mechanisms  Informs designers’ decision-making: whether or not it is desirable to deploy agents in an open society  Conflict resolution, auditing  Informs agents’ decision-making: eg when to trust another agent  Formalisation of further concepts like representation, mandate, liability, …