Hypervolume-based multi-objective reinforcement learning: Interactive approach

Abstract

In this paper, we propose a procedure of interactive multi-objective reinforcement learning for multi-step decision problems based on the preference of a decision maker. The proposed method is constructed based on the multi-objective reinforcement learning which is applied to multi-step multi-objective optimization problems. The existing literature related to the multi-objective reinforcement learning indicate that the Hypervolume is often effective to select an action from the Pareto optimal solutions instead of determining the weight of the evaluation for each objective. The experimental result using several benchmarks indicate that the proposed procedure of interactive multi-objective reinforcement learning can discover a certain action which is preferred by the decision maker through interactive.