Built a deep reinforcement learning agent capable of playing Connect 4 for a Kaggle competition.
The objective of the competition is to create an agent capable of playing Connect 4. Submissions are pinned against each other in a rolling leaderboard format.
Using a deep Q-network and experience replay, my agent is trained with Kaggle’s random and negamax agents before utilising self-play. After 10,000 games of Connect 4, my agent wins around 85% of games. For real-time updates on my agent’s performance, see Kaggle’s Connect X leaderboard. I experimented with curriculm learning (starting with smaller boards) in the hope of improving the agent’s performence, but this had little positive impact.