딥 러닝 기반의 행동 간 유사도 예측을 통한 연속 행동 공간에서의 최적 행동 탐색

Abstract

Many real-world planning problems require choosing actions in a continuous action space. The continuous action space is so large that it is often discretized using domain knowledge. However, discretization of action space causes loss of information. There is a kernel based MCTS algorithm that solves this problem through similarity between actions and finds a good action in the continuous action space. However, the algorithm does not take into account the current state and the result of taking action. This problem leads the quite inefficient action search. In this paper, we use deep neural networks to solve these problems and propose a more efficient action search algorithm.