Evaluating the Impact of Reputation-Based Agents in Social Coalition Formation

Abstract

This paper proposes a dynamic and decentralized model for coalitional skill games (CSG). The model calculates and exploits the reputation of individuals connected by a network, as an alternative to the usual CSG approaches that require reward analysis for every possible coalition to determine an optimal coalition structure for maximizing the total reward from the community. In this study, we restrict the search space for partnerships to the social neighborhoods of agents so that the social capital is used to reach a near-optimal solution by identifying how reputation can be used to better adapt the network, with the objective of bringing together agents who are more likely to cooperate in successful coalitions. In addition, our model allows a more precise quantifying of the relevance of the agents over time in social coalition formation. Experiments with different initial network topologies show that our approach is significantly better than static networks or structure-based adaptations whenever the initial network does not fit a high degree of interconnectedness, such as in a small-world model. In all the cases, the results are statistically better than current adaptation strategies.