Exploiting Physics to Learn an Optimal Swing-Up-Strategy for a Variable-Length Pendulum Using Deep Reinforcement Learning

This paper demonstrates the usage of Deep Reinforcement Learning to learn an optimal swing-up-strategy for a pneumatically actuated variable-length pendulum. For this purpose, the model-free algorithm Deep-Q-Networks is applied inside a virtual environment. Here, the influence of different reward strategies on the training success is analysed, separating between intuitive and more sophisticated functions that use pre-existing domain knowledge. The impact of the used reward function was found to be significant as the agent’s performance could be improved tremendously by inducing pre-existing knowledge about physical effects into the reward strategy.