Online Support Vector Regression based Value Function Approximation for Reinforcement Learning

Abstract

This paper proposes the online Support Vector Regression (SVR) based value function approximation method for Reinforcement Learning (RL). This approach conserves the Support Vector Machine (SVM)'s good property, the generalization which is a key issue of function approximation. Online SVR can do incremental learning and automatically track variation of environment with time-varying characteristics. Using the online SVR, we can obtain the fast and good estimation of value function and achieve RL objective efficiently. Throughout simulation tests, the feasibility and usefulness of the proposed approach is demonstrated by comparison with SARSA and Q-learning.