Evolutionary Optimization of Neural Networks for Reinforcement Learning Algorithms

Abstract

In this paper we study the combination of two powerful approaches, evolutionary topology optimization (ENZO) and temporal difference learning (TD(λ)) which is up to our knowledge the first time. Temporal difference learning was proven to be a well suited technique for learning strategies for solving reinforcement problems based on neural network models, whereas evolutionary topology optimization is concurrently the most efficient network optimization technique. On two benchmarks, a labyrinth problem and the game Nine Men’s Morris, the power of the approach is demonstrated. We conclude that this combination of evolution and reinforcement learning algorithms is a suitable framework that uses the advantages of both methods leading to small and high performing networks for reinforcement problems.