Skip to content

Latest commit

 

History

History
27 lines (23 loc) · 759 Bytes

README.md

File metadata and controls

27 lines (23 loc) · 759 Bytes

DQN agent for VizDoom

This repo consists of various DQN agents trained for different scenarios of VizDoom.

Project Structure

  • dqn_common: Common classes and functions for DQN, including DQNAgent, GameWrapper, ReplayBuffer, and process_frame
  • scenarios: scenario files of VizDoom
  • simpler_basic: DQN agent trained for simpler_basic.cfg
  • deadly_corridor: DQN agent trained for deadly_corridor.cfg

Training & Evaluation

To train an agent for a specific scenario, cd into corresponding folder and run

python3 train.py

Similarly, to evaluate a trained agent, cd into that folder and run

python3 evaluate.py

For each agent, you can tweak the model architecture in model.py and parameters/hyper-parameters in params.py.