Arnold

Why we included this project

Arnold is a PyTorch implementation of the agent from the paper 'Playing FPS Games with Deep Reinforcement Learning', which won the 2017 ViZDoom AI Competition. That pedigree matters because the code is a real competitive agent, not a simplified classroom exercise. The repo ships training scripts for DQN and DRQN variants, 17 maps across scenarios like deathmatch, defend-the-center, and health gathering, and five pretrained models you can watch or play against. The command line exposes the knobs that actually matter, replay memory size, epsilon decay, dueling networks, recurrent layers, and auxiliary game-feature prediction, so you can see how each choice changes an agent's behavior. It is aimed at researchers and students learning RL, though the code dates from 2017 and expects older PyTorch and ViZDoom versions.