Challenge a neural network trained with Deep Q-Learning. Observe live Q-value score estimates across all 7 columns as the agent decides its next move.
Estimated win expected value ($Q(s, a)$) evaluated for each move column:
This interactive demo runs a PyTorch-trained Deep Q-Network (DQN) model for Connect 4. Given a $6 \times 7$ board state tensor, the neural network evaluates action-value pairs $Q(s, a)$ for all 7 columns simultaneously.
Integrates Dueling Networks (separating state value $V(s)$ and advantage $A(s, a)$), Double Q-learning (mitigating overestimation bias), and Prioritized Experience Replay (PER).
Trained through millions of self-play episodes in PettingZoo multi-agent environments, progressively discovering tactical tactics (forking, double traps, and defensive block prioritization).