CrashCourse AI episode 12
Posted on: November 3, 2019
Why use games with AI? Games provide training data. An AI can play games or learn from another entity playing games.
Games are so good at training AI with strategies that AI beat human opponents at:
- Chess in 1997
- Go in 2017
- DOTA 2 in 2018
- Starcraft 2 in 2019
Whatever the game, the central tenet seems to be the minimax algorithm.
But the flaw on relying only on this algorithm is the premise that you can work out every strategic possibility. As the video explained, the combinations and permutations could outnumber the number of atoms in the known universe. This is not practical or feasible. Instead, an AI might estimate the chances of winning from a smaller set of possibilities.
AI can learn from games and even beat humans because there are clear rules. Where it currently fails to keep up are in areas like humour, social cues, creativity, and surprise.
Leave a Reply