This repository contains a Deep Q-Network (DQN) agent trained to play multiple Atari games, achieving an average reward of +15.3 compared to a baseline of -21. The agent was trained using Python and Jupyter Notebook, with a training speed of 420 steps per second and convergence at episode 350.
From the README
Deep Q-Network (DQN) for Atari Pong Quick Results Summary Final Performance: +15.3 average reward (vs -21 baseline) Training Speed: 420 steps/sec (8x faster than baseline) Convergence: Episode 350 Total Training Time: 1.2 hours (500 episodes) DQNHyperparams: • algo: dqn • env: ALE/Pong-v5 • device: auto • batchsize: 32 • buffersize: 100000 • explorationfinaleps: 0.0099 • gradientsteps: 2 • learningrate: 0.0001 • discountfactor : 0.99 • learningstarts: 100000 • targetupdateinterval: 1000 • trainfreq: 8 • envhyperparams: • framestack: 4 • nenvs: 8 Training Progression Key Milestones • Episode 100: -18.2 avg (still mostly random) • Episode 200: -12.5 avg (learning to track ball) • Episode...
Skills
- Python
- Jupyter Notebook