I wrote a small AGI project based on Eric Baum's clever Hayek Machine using the game of Go as a test bed.

Go provides an excellent compromise between the bumpy landscape of simpler toy problems and the real world. Go is rich, clean, concrete, relatively simple to implement and provides a direct measure of intelligence via its well defined ranking system.

I took a strict artificial general intelligence approach where no code was given to the AGI for rules, strategy, evaluation function, min-max, etc. The AGI had to discover everything.

The AGI was able to win only against a very simple player - though by 16 points on a 4x4 board and a penalty of 5 points and turn forfeiture for illegal moves. The simple player plays random moves, knows which moves are legal and knows to avoid self capture.

The program played approx. 2,000 4x4 games per second on a conventional PC.

I experimented with conventional assembly-language-like instruction sets as well as Markov algorithms as the core evolutionary programming substrate. I experimented with such variations as a recursive version of Hayek where agents can be Hayeks and assorted other modifications.

The conventional instruction set performed better than Markov algorithms and the non-recursive version worked better than the recursive version. I have not tried the Post system used by Baum as it did not appear as well suited to Go and runs relatively slowly.

There may be ways to adapt the Markov algorithm to 2D space, as applied it to a 1D algorithmic representation of an "eye" movement path.

From this project my impression is that ideas in Hayek need to be combined into a more sophisticated system to create substantial intelligence.

Posted 8/5/07, Updated 9/29/07