Perpetually playing physics

Abstract

Here we discuss ideas of reinforcement learning and the importance of various aspects of it. We show how reinforcement learning methods based on genetic algorithms can be used to reproduce thermodynamic cycles without prior knowledge of physics. To show this, we introduce an environment that models a simple heat engine. With this, we are able to optimize a neural network based policy to maximize the thermal efficiency for different cases. Using a series of restricted action sets in this environment, our policy was able to reproduce three known thermodynamic cycles. We also introduce an irreversible action, creating an unknown thermodynamic cycle that the agent helps discover, showing how reinforcement learning can find solutions to new problems. We also discuss shortcomings of the method used, the importance of understanding the class of problem being handled, and why some methods can only be used for certain classes of problems.