Video games have recently been used to train and test artificial-intelligence (AI) algorithms that learn to behave effectively in a wide array of environments. Deep RL approaches reach good asymptotic performance on certain classes of video games. But unlike humans, they require large amounts of data and generalize poorly to even slight changes to their environments. This project aims to explore the idea that having a human-like capacity to learn requires building a closely-fitting model of the world. The specific aim of this project is to develop a language that is sufficiently rich and flexible to be able to describe a wide range of possible game-worlds and to develop inference procedures for learning a good game description given actual game-play data.

This SuperUROP will be a continuation of the work I have been doing this summer with the lab, where we have been developing this video-game description language. While the UROP gave me the chance to develop a wide range of skills in computer science, this SuperUROP project will allow me to focus on the really interesting and more challenging aspects of AI and reinforce the learning I have been wanting to do since joining this lab.