Modeling multi-objectivization mechanism in multi-agent domain

Abstract

Many real-world tasks require making sequential decisions that involve multiple conflicting objectives. Furthermore, there exist multiple decision-makers, called multiagent, each of whom pursues its own profit. Thus, each agent should take into account the effect of other agents ‘ decisions to reach a point of compromise. For example, each agent decides with thought of other agents ‘ behavior in the decision of selecting the faster driving route to the destination, selecting a supermarket checkout line, and so on. For solving a sequential multi-objective decision problem, a multi-objective reinforcement learning (MORL) approach has been investigated.However, current research on MORL cannot deal with the multi-agent system where existing agents are influenced one another. Therefore, in this study, we expand the conventional multi-objective reinforcement learning by introducing the idea of multi-objectivization with dynamic weight setting of other decision-makers. In an experiment, our proposed model with dynamic weight can express the cooperative behaviors that seems to be considered other decision-makers in the multiagent environment.