The first entry for this month, I mention the three main competing interests for my energies and attention: the day job, writing, and the neural net project.

As it happens, most of my energy has been directed toward this last. It's where my thoughts keep drifting back to, and it's what has interested me most lately. I'm not sure this is a bad thing. A much worse thing would be for me to have no intellectual life, of course. As you may have noticed here, for example, I've done a lot less wrangling over political issues, wanting much more to spend those brain cycles puzzling over difficult and interesting problems.

So the past couple of weeks I've been reading quite a bit of the academic literature on genetic algorithms applied to game domains, from Tic-Tac-Toe to checkers to Texas Hold'em Poker.

Philip and I are still a ways off from tackling a game domain (Tic-Tac-Toe will be our first) but I think the issues of how to set up the population and have the individuals compete is extremely interesting.

Here are a few approaches:

--In the domain of Backgammon, one researcher basically cloned each neural net player and had it play against itself. The connections were strengthened for the winning clone and weakened for the loser, using a training method called Temporal Difference. TD-Gammon, the player trained this way, is one of the strongest Backgammon AIs around.

--Another researcher, in Checkers, created a population of 30 neural nets. Each generation, a given net would play 5 other random opponents from the population, accumulating points for wins and draws. The 15 top players each generation would survive and spawn offspring. The other 15 would die. Blondie24, the Checkers program evolved this way, has a rating over 1900 (1600 is an average player).

But I'm particularly interested in those experiments which tried to exploit coevolution, in which two populations are created and compete against one another. The desired effect is to create an "arms race" between the populations. This kind of competitive coevolution can be seen throughout nature. For example, cheetahs and antelope. Each gets faster because the slower individuals from each population die off.

The desire is to create the same sort of upward spiral of competition between two populations playing each other in the same game domain. I've got some ideas, and hope to be able to implement them soon.