The Adaptive Multi-personality Agent

Negotiating agents are an increasing factor in large-scale multi agents systems today. Thus, the need to design a high-performance agent, able to interact with its surrounding in order to achieve its goals, is in demand. It is especially beneficial to design an agent able to perform well in all environments a cooperative environment where all agents work together to achieve a joint goal (for example RoboCup), a competitive environment where each agent has its own set of goals and competes with the other agents in the system (for example auctions) or an intermediate environment where agents have their own set of goals but also share a joint one (for example branches of the same company working to maximize their own income, while keeping the company interests in mind as well). Throughout the years many models of agents have been developed, most of which have been designated to act in a specific type of environment, either a cooperative or a competitive environment. Furthermore, several techniques for dealing with agents that compromise the foundations of the cooperative/competitive environment were developed, in order to increase the agents’ gain and/or protect them from exploiters. Our solution is to design an Adaptive Multi-personality agent that consists of a set of sub-agents. Each sub-agent is in charge of interacting and negotiating with one of the other agents coexisting in its surrounding. By modeling agents it interacts with, and by learning which sub-agent best-suits each agent, the Adaptive Multi-personality agent can interact with other agents in an optimal manner. As a result, the Adaptive Multi-personality agent is able to perform well in all types of environments, cooperative, competitive and intermediate alike. Moreover, it copes well with different strategies agents deploy it doesn’t yield to exploiters while taking advantage of the selfless. The first part of this thesis describes the Adaptive Multi-personality agent: its design and motivation, its different modules, its special instance as an adaptive one-personality agent and our hypotheses regarding its performance. The second part introduces the domain in which we would evaluate the Adaptive Multipersonality performance. This domain is the Colored Trails game, a complex game with numerous parameters, which allows us to conduct a large number of different experiments. In this part we also discuss the additional agents we have at our disposal, which were designed by other designers, and will be used to evaluate the Adaptive Multi-personality performance as well. The last part presents the experiments we executed, some of which were precursory experiments, which “trained” the agent, and the others were evaluation experiments. All in all we executed over 58,000 games, which translate into ~3800 hours of computation. In all those experiments the Adaptive Multi-personality agent proved to be significantly better than its adaptive one-personality instance and the additional agents alike, and reached higher scores.