[4E1-GS-2] Machine learning: agents

[4E1-GS-2-04] Calculating weight of objectives to find complete pareto front for multi-objective reinforcement learning

〇Hirotaka Naiji¹, Sachiyo Arai² (1. Department of Electrical and Electronic Engineering, Graduate School of Engineering, Chiba University, 2. Department of Urban Environment Systems, Graduate School of Science and Engineering, Chiba University)

Keywords:reinforcement learning, decision making

In multi-objective reinforcement learning, there are trade-off solutions. In order to choose a desirable solution among them, we need to find weights that are decision criteria. Existing methods assume that the weights are uniform at each stage from the initial state to the final state. As a result, it seems that Pareto-optimal policies that cannot be obtained will occur. A method is proposed to obtain Pareto-optimal measures, which cannot be obtained by existing methods, by extracting the policies considering that the weights change at each stage of decision making. However, since the weights are searched exhaustively when extracting Pareto policies, computational complexity is an issue. In this paper, we attempted to reduce the computational complexity by efficiently searching the weights of Pareto optimal policies. Through experiments, we verify that the weights can be estimated and reduce the computational complexity compared to the existing method.