#VoxxedDaysCERN26
AI has become part of our everyday lives but we often forget one of its earliest and most fascinating uses: creating an opponent to play games with.

We’ve seen it in chess, Go, and countless video games, where AI opponents follow complex strategies or fixed rule sets. But who says you need PyTorch or TensorFlow to do it? Sometimes, the good old Java is all you need.

In this talk, I’ll share how I built a Deep Q-Network (DQN) in Java to teach an AI to play the board game Azul. I’ll cover environment design, state encoding, reward shaping, and training strategies.

Devoxx

A full-stack engineer at CERN presents a personal project implementing Deep Q-Network (DQN) AI in Java to play the board game Azul. The talk covers the game rules, Q-learning fundamentals, the transition from Q-tables to neural networks for large state spaces, and a dueling DQN approach where two agents train against each other. Implementation details include state encoding (factories, player boards, walls, scores), epsilon-greedy exploration policy, replay buffer, He weight initialization, and a reward system based on score deltas with floor-line penalties. A live demo shows the trained AI making reasonable moves against a human player.

Training My Rival in Java: A Deep Q-Learning AI to Play Azul by Victor Uria Valle