A Genetic Algorithm for Automatically Designing Modular Reinforcement Learning Agents

Reinforcement learning (RL) is one of the machine learning techniques and has been received much attention as a new self-adaptive controller for various systems. The RL agent autonomously learns suitable policies via the clue of reinforcement signals by trial and error. Most RL methods have a serious problem that computational resources and time to learn appropriate policies grow rapidly as the size of the problem space increases. To overcome the problem, the Modular Reinforcement Learning Architecture (MRLA) has been proposed and shown good results. However, the performance of an agent with MRLA deteriorates unless the agent consists of appropriate modules for a given task, which cannot be predicted in advance. This means that a human designer has to identify a good combination of modules by trial and error. In this paper, we propose a genetic algorithm for finding appropriate combination of modules and show its effectiveness by applying it to a benchmark problem.

• Type: Conference
• Year: 2000
• Where: GECCO
• Authors: Isao Ono, Tetsuo Nijo, Norihiko Ono
• Keywords: GECCO 2000, Most RL Methods, Optimization, Reinforcement Learning, RL Agent