강화학습 상태 텐서 차원에 따른 학습 효과도

Abstract

In reinforcement learning, an agent takes actions in an environment, which is interpreted into a reward and a “representation of the state”. It is well known that the performance of the rein forcement learning is dependent on the “data model representing the state” of a given environ ment. This paper proposes a data model representing the state which is suitable for a FPS (First Person Shooting) game, a military tactics simulator that changes state extremely and needs deci sion making quickly. The proposed data model consists of matrix (multi-dimensional tensors) for spatial features and vectors for non-spatial features. To prove the usefulness of the proposed data model, this paper shows experimental results for a FPS game.