This work sought to determine if deep Q learning (DQN) can outperform classical control techniques in the context of controlling the behavior of fast-slow dynamical systems. These methods were applied to the cart-pole system and acrobot system. Through the analysis, we found that reinforcement learning controllers can outperform dynamics based controllers, specifically when there are many nodes per layer in the DQN.
Deep Q Learning Control Approaches
Dynamic Control Approaches
Training Reinforcement Learning of Car-Pole
Completed RL Controller for the Cart-Pole System
Completed RL Controller for Acrobot System
This website offers a brief overview of what we accomplished in the spring semester of 2018. Please don’t hesitate to reach out to any of us with questions or comments about our work.