Training My Rival in Java: A Deep Q-Learning AI to Play Azul
summary
Built a Deep Q-Network in Java (no PyTorch or TensorFlow) that encodes Azul state and actions (factories or center, color, target row, wall, pattern lines, tile bag size) and trains with a replay buffer and target network. Uses epsilon-greedy 0.8→decay exploration, gamma 0.99 discount, reward = round score delta minus floor penalties. Checkpoints save every 1,000 games to training.bin and the AI plays reasonably against a human in the demo.
AI-written summary · about
chapters (8)
- 0:00Azul AI project intro
- 1:04Azul game rules and implementation
- 5:00Q-learning and neural basics
- 7:27Dueling DQN architecture
- 9:37DQN agent and hyperparameters
- 12:04State encoding and action policy
- 13:55Training process and rewards
- 15:24Live demo: human vs AI
related talks
21mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
56mCan I do AI?
35mLet's play Space Invaders!
49mML: Teaching machine to learn from its own mistakes (...)
33mBreakthroughs and the Future of (Deep) Reinforcement Learning
01h 01mReinforcement Learning: Pac-Man
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
37mHow I trained an AI Model to Beat the 1990's Arcade Game Double Dragon