RL for game playing

Games are RL's favourite proving ground — they have crisp rules, a clear score, and unlimited cheap practice, so an agent can play millions of matches and measure its progress exactly. The landmark results of the field, from Atari to Go to StarCraft and Dota, all came from games.

Different games stress different skills. Atari from pixels demands perception and credit assignment; board games demand long-horizon planning and search; real-time strategy games pile on partial observability, huge action spaces, and team coordination. The methods range from value-based learning (DQN) through policy gradients and actor-critic, often combined with self-play and tree search.

Mastery of a game does not automatically transfer to messy real tasks. Games hand the agent a perfect simulator, dense and immediate feedback, and a single fixed objective — three luxuries the real world almost never provides.