Hacking NetHack: Novel Reinforcement Learning Architectures for Multi-Objective Optimization

Despite significant progress in the field, designing reinforcement learning algorithms for stochastic, high dimensional, and reward sparse environments remains difficult and computationally expensive. The roguelike game NetHack provides an opportunity to explore these open problems with minimal resources for simulation, and has recently been packaged as a model system for reinforcement learning tasks. In our project, we worked to extend the capabilities of the NetHack Learning Environment and investigated the learning efficiency of deep reinforcement learning algorithms. Although we found some strategies for improving learning to be promising, high variability across trials prevented us from reaching conclusive results. We have released our code and extensions to the NetHack Learning Environment at https://github.com/apandit42/cs230-deep-hack