Benchmarking genetic programming in a multi-action reinforcement learning locomotion task
Ryan Amaral, Alexandru Ianta, Caleidgh Bayer, Robert J. Smith, Malcolm Iain Heywood · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
Reinforcement learning (RL) requires an agent to interact with an environment to maximize the cumulative rather than the immediate reward. Recently, there as been a significant growth in the availability of scalable RL tasks, e.g. OpenAI gym. However, most benchmarking studies concentrate on RL solutions based on some form of deep learning. In this work, we benchmark a family of linear genetic programming based approaches to the 2-d biped walker problem. The biped walker is an example of a RL environment described in terms of a multi-dimensional, real-valued 24-d input and 4-d action space. Specific recommendations are made regarding mechanisms to adopt that are able to consistently produce solutions, in this case using transfer from periodic restarts.