Abuudiii/dqn-QWOP

DQN agent playing QWOP

A Deep Q-Network agent trained with Stable-Baselines3 on qwop-python, a headless Box2D recreation of Bennett Foddy's QWOP.

• Actions: 16 Q/W/O/P key combinations, each held for 4 frames
• Observations: 60 numbers (body-part positions, angles and velocities)
• Training: 1M steps on an AMD Radeon AI PRO R9700 (ROCm)
• Result (deterministic, 10 episodes): mean reward 178.7, runs 8.8 m in 14.3 s before falling

Hyperparameters: see train_dqn.yml (changed from the qwop-python defaults: total_timesteps: 1000000, target_update_interval: 10000, exploration_fraction: 0.3) and env.yml.

Usage

from stable_baselines3 import DQN
model = DQN.load("dqn-QWOP.zip")