Heuristically Accelerated Reinforcement Learning: Theoretical and Experimental Results

摘要

Since finding control policies using Reinforcement Learning (RL) can be very time consuming, in recent years several authors have investigated how to speed up RL algorithms by making improved action selections based on heuristics. In this work we present new theoretical results - convergence and a superior limit for value estimation errors - for the class that encompasses all heuristics-based algorithms, called Heuristically Accelerated Reinforcement Learning. We also expand this new class by proposing three new algorithms, the Heuristically Accelerated Q(A), SARSA(A) and TD(A), the first algorithms that uses both heuristics and eligibility traces. Empirical evaluations were conducted in traditional control problems and results show that using heuristics significantly enhances the performance of the learning process.