Chapter 2. Go as a machine-learning problem
This chapter covers
- Why are games a good subject for AI?
- Why is Go a good problem for deep learning?
- What are the rules of Go?
- What aspects of game playing can you solve with machine learning?
Games are a favorite subject for AI research, and it’s not just because they’re fun. They also simplify some of the complexities of real life, so you can focus on the algorithms you’re studying.
Imagine you see a comment on Twitter or Facebook: something like, “Ugh, I forgot my umbrella.” You’d quickly conclude that your friend got caught out in the rain. But that information isn’t included anywhere in the sentence. How did you reach that conclusion? First, you applied common knowledge about what umbrellas are for. Second, you applied social knowledge about the kinds of comments people bother to make: it’d be strange to say, “I forgot my umbrella” on a bright, sunny day.
As humans, we effortlessly factor in all this context when reading a sentence. This isn’t so easy for computers. Modern deep-learning techniques are effective at processing the information you supply them. But you’re limited in your ability to find all the relevant information and feed it to computers. Games sidestep that problem. They take place in an artificial universe, where all the information you need in order to make a decision is spelled out in the rules.