News for hackers. Stuff that matters.
Re:A 1979 War-Game That Takes 1,500 Hours to Complete by gwern on Saturday September 23, 2017 @05:17PM (#15321778)
Random sampling wouldn't work because it is the essence of cheese/game-breaking to go into rare or weird edge-cases and parts of state-space. Imagine a broken fighting game where you could win by simply executing 'upper block' the entire game; random sampling will be exponentially unlikely to sample a game with all upper-blocks. But if you used a mechanism like self-play reinforcement learning to explore the game state-space, your agents might autonomously discover game-breaking strategies on their own.

There are no comments on this comment yet.