Unified Algorithm to Improve Reinforcement Learning in Dynamic Environments - An Instance-based Approach

This paper presents an approach for speeding up the convergence of adaptive intelligent agents using reinforcement learning algorithms. Speeding up the learning of an intelligent agent is a complex task since the choice of inadequate updating techniques may cause delays in the learning process or even induce an unexpected acceleration that causes the agent to converge to a non-satisfactory policy. We have developed a technique for estimating policies which combines instance-based learning and reinforcement learning algorithms in Markovian environments. Experimental results in dynamic environments of different dimensions have shown that the proposed technique is able to speed up the convergence of the agents while achieving optimal action policies, avoiding problems of classical reinforcement learning approaches.