This repository contains the implementation of a simple collision avoidance game which can be played by humans or software-defined agents. It also includes four agents: a Bernoulli random agent, a discretized Q-fit agent, a neural Q-fit agent with a dense network architecture, and the convolutional Q-fit agent from "Human-level control through deep reinforcement learning" (Mnih et al, 2014). There is a simple API for training, evaluating, and saving agents to disk, but the code is experimental, terse, and not very well structured. I was able to train a dense agent with superhuman performance at the collision avoidance task, but my current hardware makes it difficult to train a comparable convolutional agent in a reasonable amount of time.

A full description of the motivation, experimental setup, and results is available in the pdf.