FirstFifty wrote:The program generates a capturing race puzzle, the player picks a color, then plays it out against the AI. The board configurations are unnatural looking.
If I understand what you're saying, then this approach won't work. Or, at least, it won't work the way I think it should!
If the program generates the puzzle randomly and lets the player choose the puzzle, then you need to be assuming that the puzzles are always equally hard for both sides, and the bot is always just slightly dumber than the human.
First of all, playing a random capturing race against an extremely dumb bot isn't going to reveal any of the subtlety of go, right? Second, there's no way to fine-tune the AI such that the rate of progress is going to increase at the exact same rate that the person doing the puzzles improves. What you really want, I think, is to figure out a sequence of puzzles - always biased in favor of Black - that a beginner can start out with and find the first one extremely intuitive, even simplistic, but find the final ones nearly impossible. You could design the puzzles by hand and have just a few possible sequences (this would, I think, be easiest) or you could come up with the basic principles that make capturing races easier and harder and have the program generate random puzzles that embody some mix of principles that generate more and less complex races that are more or less tilted in favor of Black (depending on how far the player has advanced). Then let Fuego take White and watch the sparks fly.
