07 / Learning game
Laser Battle 2
Watch the saved DQN policies, then try the arena controls yourself.
DQN checkpoint replay · playable scripted duel
07 / Recorded self-play & controls
Two policies in the arena.
The checkpoint replay uses the original environment, obstacles, ammunition, and greedy DQN actions. You can also try the browser duel.
1000_ep_fix_v2 · champion_final vs challenger_final · seed 42, three consecutive maps.
Move with WASD or arrows; aim with the pointer; fire with Space or click. On mobile, drag to move and use Fire.
Interactive browser sketch against a scripted opponent.
The project
Self-play with a champion
A compact DQN challenger learns against a champion in a two-dimensional arena. The original environment includes turning, firing, reloading, health, and random obstacles with a traversable path between spawns. The selected run trained for 1,000 episodes and promoted the challenger nine times.
Observations & lessons
What the saved policies show
Both final greedy policies draw on these three consecutive maps, with no health lost. Promotions during exploratory training did not guarantee an effective deterministic duel. The reward and win-rate plots help locate that gap between the training schedule and checkpoint behavior.
Recorded plots
The self-play training run
