The True Online Continuous Learning Automation (TOCLA) in a continuous control benchmarking of actor-critic algorithms

摘要

Reinforcement learning problems are often discretised, use linear function approximation, or perform batch updates. However, many applications that can benefit from reinforcement learning contain continuous variables and are inherently non-linear, for example, the control of aerospace or maritime robotic vehicles. Recent work has brought focus onto online temporal difference methods, specifically for using non-linear function approximation. In this paper, we evaluate the Forward Actor-Critic against the regular Actor-Critic, and Continuous Actor-Critic Learning Automation. We also propose and evaluate a new algorithm called True Online Continuous Learning Automation (TOCLA) which combines these two approaches. The chosen benchmark problem was the Mountain Car Continuous-vO environment from OpenAI Gym, which represents a further step in complexity over the benchmark used to test the Forward Actor Critic in previous works. Our results demonstrate the superiority of TOCLA in terms of its sensitivity to hyper-parameter selection compared with the Forward Actor Critic, Continuous Actor Critic Learning Automation, and Actor Critic algorithms.