Bargaining in a Three-Agent CoalitionS Game: An Application of Genetic Programming

We are conducting a series of investigations whose primary objective is to demonstrate that boundedly rational agents, operating with fairly elementary computational mechanisms, can adapt to achieve approximately optimal strategies for bargaining with other agents in complex and dynamic environments of multilateral negotiations that humans find challenging. In this paper, we present results from an application of genetic programming (Koza, 1992) to model the co-evolution of simple artificial agents negotiating coalition agreements in a threeagent cooperative game.2 The following sections summarize part of the scientific literature that motivates this research, describe briefly the genetic programming approach we use to model this game, then present results demonstrating that, through a process of co-evolution, these artificial agents adapt to formulate strategies that cope reasonably well under difficult circumstances to negotiate coalition agreements that not only rival those achieved by human subjects but also approximate those prescribed by cooperative game theory as the solution of this game.