関数近似手法を用いた強化学習アルゴリズム

概要

論文の詳細を見る
In this paper, we propose a new RL algorithm with Locally Weighted Partial Least Squares (LWPLS) as a function approximator. LWPLS is a class of techniques from nonparametric statistics that is ideally suited to reduce the computational complexity and to avoid numerical problems. The principle of LWPLS is to fit linear models using a hierarchy of univariate regressions along selected projections in input space. The projections are chosen according to the correlation between input and output data, and the algorithm assures that subsequent projections are orthogonal in input space. This new RL algorithm is compared with the usual way of quantizing the state space with grids in a mobile robot navigation task. The results of the extensive simulation demonstrate that our proposed algorithm is clearly outperforming the usual way.
八戸工業高等専門学校の論文