Learning To Identify Beneficial Partners

Human and artificial agents routinely make critical choices about interaction partners. The decision about which of several possible candidates to interact with, either for a limited or extended time period, has significant importance on the competitiveness, survivability, and overall utility of an agent. We assume that an agent has time and resource constraints that limit its participation to only a fixed number, k, of relationships or interactions with other agents in a particular time period. Therefore, in a given time period, an agent is free to choose to interact with any k other agents from a society of N agents. A bilateral relationship is established in a time period, however, if both agents choose to do so. The goal of this research is to investigate the extent to which known learning schemes can identify and sustain mutually beneficial relationships in these conditions. While exploration is necessary to locate possible fruitful relationships, resource constraints limit the extent of exploration. The desired emergent phenomena of mutual cooperation is uncertain and fragile as it is predicated on the convergence of the learning of multiple, concurrent learners. We investigate the success of individual learners in identifying and sustaining mutually beneficial relationships in a multiagent society under varying envrionmental conditions.