Evolution of a Go program

About the development of Moyo Go Studio, software to (help) play the Oriental game of Go. Go is a two-player zero-sum game of perfect information. It is considered much harder than Chess. Currently, in spite of enormous effort expended, no computer program plays it above the level of a beginner.

Wednesday, October 25, 2006

Bone For Microsoft

The boffins at Microsoft appear to be stuck with their Bayesian network after they copied it from me, and they have been trying to pick my brain using nefarious means.

Here-ya-go:

Interview with Jeff Hawkins on his book "On Intelligence".

Paper by Jeff Hawkins on Hierarchical Temporal Memory.

Powerpoint presentation for his talk at IBM's Almaden Institute.

..and the lecture itself:




HTM's are currently perhaps the best way to make a Go program using a neural network. A game of Go is a temporal "world" with "causes" (chains, order-liberties, patterns etc.) and a HTM is able to make predictions (finding good moves, in HTM-terms: "beliefs"). It learns pattern sequences over time and is much more sophisticated and powerful than any other "ordinary" ANN. So it would be a logical evolutionary step for Microsoft's Bayesian pattern network.

Finally, a link to his book.