Reinforcement Learning and Trajectory Planning based on Model Approximation with Neural Networks applied to Transition Problems

Abstract

Abstract In this paper we use a multilayer neural network to approximate the dynamics of nonlinear (mechanical) control systems. Furthermore, these neural network models are combined with offline trajectory planning, to form a model-based reinforcement learning (RL) algorithm, suitable for transition problems of nonlinear dynamical systems. We evaluate the algorithm on the swing-up of the cart-pole benchmark system and observe a significant performance gain in terms of data efficiency compared to a state-of-the-art model-free RL method (Deep Deterministic Policy Gradient (DDPG)). Additionally, we present first experimental results on a cart-triple-pole system test bench. For a simple transition problem, the proposed algorithm shows a good controller performance.