RL Agent for Tron
Light cycles, learned
- 74.4%win rate against a random player
- 41.9%wins against survival strategies (19.3% draws)
I trained a Proximal Policy Optimization agent to play 1v1 Tron competitively, starting with behavior cloning and then curriculum learning against increasingly strong BFS and Voronoi agents.
To measure it I built a 3,000-game evaluation harness. The agent won 74.4% of games against a random player and 41.9% against survival strategies, with 19.3% draws.