This repo contains deep reinforcement learning methods applied to the VizDoom environment. Of the VizDoom scenarios, we chose to evaluate agent performance on the Defend the Line and Deadly Corridor scenarios. More detail on how the environments work can be found in the VizDoom documentation linked above.
We tested the following methods:
- DQN: see
agents/dqn/README.md - Double Dueling DQN with Prioritised Experience Replay (which we have referred to as Improved DQN): see
agents/improved_dqn/README.md - PPO: see
agents/ppo/README.md
We recommend creating a Python virtual environment. Then, install project dependencies using pip install -r requirements.txt.
We have used agents that make completely random actions as baselines - this script used to record these mean returns can be found in random-baselines/, which can take an environment and a given number of episodes to average over to find a mean episodic return value for random action performance.