The Taub Faculty of Computer Science Events and Talks

Classic reinforcement learning methods such as Q-learning and policy gradient methods have seen great success in learning to perform a plethora of common tasks in AI, from playing video games to controlling self-driving vehicles and beyond. As a result, these methods and their extensions have become standard in most reinforcement learning settings. However, they have trouble adapting to changes in the environment. We believe this issue can be solved by giving the agent awareness of these changes in the form of context that could potentially steer it toward an optimal policy regardless of the nature of these changes. In this proposal, we suggest a novel learning method that provides said context in the form of a graph abstraction of the current environment to the agent. We do this by modeling the contexts as reward machines and using graph neural networks to embed them into meaningful representations that an agent can leverage to understand the current environment status.