Reinforcement Learning (RL) is a well-known technique for the solution of problems where agents need to act with success in an unknown environment, learning through trial and error. However, this technique is not efficient enough to be used in applications with real world demands due to the time that the agent needs to learn. This paper investigates the use of Transfer Learning (TL) between agents to speed up the well-known Q-learning Reinforcement Learning algorithm. The new approach presented here allows the use of cases in a case base as heuristics to speed up the Q-learning algorithm, combining Case-Based Reasoning (CBR) and Heuristically Accelerated Reinforcement Learning (HARL) techniques. A set of empirical evaluations were conducted in the Mountain Car Problem Domain, where the actions learned during the solution of the 2D version of the problem can be used to speed up the learning of the policies for its 3D version. The experiments were made comparing the Q-learning Reinforcement Learning algorithm, the HAQL Heuristic Accelerated Reinforcement Learning (HARL) algorithm and the TL-HAQL algorithm, proposed here. The results show that the use of a case-base for transfer learning can lead to a significant improvement in the performance of the agent, making it learn faster than using either RL or HARL methods alone.
17 February 2020 in Optimal Control Applications and Methods
In this paper we propose to combine Case-based Reasoning and Heuristically Accelerated Reinforcement Learning to speed up a Reinforcement Learning algorithm in a Transfer Learning problem. To do so, we propose a new algorithm called SARSA Accelerated by Transfer Learning - SATL, which uses Reinforcement Learning to learn how to perform one task, stores the policy for this problem as a case-base and then uses the learned case-base as heuristics to speed up the learning performance in a related, but different, task. A set of empirical evaluations were conducted in transferring the learning between two domains with multiple agents: an expanded version of Littman's simulated robot soccer and the RoboCup Soccer Keep away. A policy learned by one agent in the Littman's soccer is used to speed up the agent learning in the Keep away soccer. The results show that the use of this new algorithm can lead to a significant improvement in the performance of the learning agents.