Back to Posts

Why teach AIs how to play games

Given there are far more pressing problems that AIs can potentially solve, say drug discovery, road traffic optimization, stock market predictions etc, why is it that every time every AIs win a particular game, it is considered a big deal?

AIs have so far beaten human world champions at chess, go, Dota, Starcraft and now even Poker.

The reason these are big advancements is because games model life.

If you’ve ever played an immersive RPG like GTA or Witcher, you will understand what I mean. These games go to such extents that it will soon be hard to tell them apart from real life.

Games model:

  • Physics Cardboard will float but your gun will sink.
  • Social structure If you start running in a crowd, others around will notice.
  • Common sense Computer players take cover behind objects when under attack.
  • Strategy Enemies invest in fortification when you invest in artillery.

Of course, they do not model life exactly. But remember, the map is not the territory.

So any decently good approximation is a terrific playground for AI to practice. If an AI can learn the rules of a game(subset), then it can learn the rules of life(superset).

The frontier AI group OpenAI has even built a gym where you can train AI for many games if you’re interested.

Read Next

Don't teach the basics