O grupo de pesquisa em "Teoria e Inteligência Computacional" da UFRN tem o prazer de anunciar a palestra:
Towards the self-regulation of personality-based social exchange processes in multiagent systems
Prof. Graçaliz Dimuro
Universidade Católica de Pelotas
Data: 26/09/2008 às 14h
Local: UFRN, DIMAp, sala de reuniões
The regulation of agent interactions based on Piaget's theory of social exchanges was first proposed by Dimuro and Costa in 2005. According to that theory, social interactions are seen as service exchanges between pairs of agents and the qualitative evaluation of those exchanges by the agents by means of social exchange values: the investment value for performing a service or the satisfaction value for receiving it. The exchanges also generate values of debts and credits that help to guarantee the continuity of future interactions. A multiagent system (MAS) is said to be in social equilibrium if the balances of investment and satisfaction values are equilibrated for the successive exchanges occurring along the time. A (centralized) mechanism for the regulation of social exchanges in MAS, based on the concept of equilibrium supervisor with an associated Qualitative Interval Markov Decision Process (QI-MDP) was introduced in 2005. This approach was extended to consider personality-based agents in 2006.
We are now going towards the (distributed) self-regulation of personality-based social exchanges, by decentralizing the equilibrium supervisor through the internalization of its decision process in the interacting agents. For simplicity, we consider that the exchanges between each pair of agents have no influence from the exchanges occurring between other pair of agents, which allowed for a simple way of distributing the decision process. However, since the agents do not necessarily have access to the internal states (balances of exchange values) of each other, the decision processes have to operate in a partially observable mode, by means of Partially Observable Markov Models POMDP).
The decision about the best exchanges that an agent should propose to its partner in order to achieve social equilibrium, or to promote new interactions, is modeled as a global POMDP for each personality trait that its partner may assume. Considering a set of six personality traits (egoism, strong egoism, altruism, strong altruism, tolerance, equilibrium fanaticism), each global POMDP is decomposed into three sub-POMDPs, according to the current internal state (favorable, equilibrated or unfavorable balance of material exchange values) of the agent that is supervising the interaction.
For convenience, we have chosen the BDI architecture for the agents in the social exchange simulator used in this work. We developed an algorithm to extract BDI plans from the policy graphs of each sub-POMDP, building a set of rules that form the BDI plans for interacting with agents of each personality model. Those plans are to be selected in each interaction according to the current balance of material values of the agents in each exchange stage. Such plans are said to ``obey'' optimal POMDP policies.