14:00 - 14:45
We introduce a case study where the goal was to improve the performance of an existing PIcontroller for a vehicle simulation on an oval circuit, using a Reinforcement Learning (RL) agent. It is an example of applying RL to solve a realistic (and complex) problem. Typically, there are two main challenges associated with a control-oriented RL problem: Finding a suitable network architecture for the RL agent, and designing an accurate reward function that can efficiently guide the agent during training and help minimize the training time, which is a well-known RL bottleneck. We will describe our solution approach in detail as well as the problems associated with training the RL agent. The results indicate that the RL agent can be used to outperform a strategy which employs only a traditional controller. In addition to this, our case study also contains an example of applying imitation learning.