ANNEvolve :: Evolution of Artificial Neural Networks

In the biological world, a creature like a cheetah is supremely adapted to hunt animals such as gazelles, which are supremely adapted to escape from animals such as cheetahs. Both are the result of millions of generations of evolution.

But, and here is my point, in the early days of the evolution of the cheetah line, they did not have to catch gazelles in order to eat. Their prey was slower then. Similarly, the early pre-gazelles did not have to escape from cheetahs in order to survive, because the pre-cheetahs were not as capable as cheetahs.

What this means to us, who are simulating evolution, is that if you want to evolve something very capable, you probably have to do it in a series of small steps, beginning with easier challenges. In the case of tic-tac-toe, the early neural nets cannot really play tic-tac-toe. First they have to evolve to make legal moves in simple board patterns. Later they learn to block 2 in a row. By a series of small steps we can evolve them.

One approach to doing that I learned from reading parts of a doctoral dissertation by Chris Rosin. His idea is to have two co-evolving populations, each of which creates the problem for the other to solve. The general idea is to have each population present an initially easy challenge to the other. As the two populations evolve, both are presented with challenges of increasing difficulty. In the case of tic-tac-toe, we might have one population which are the first-movers, and another population that are the second-movers. Initially, neither population can really play ttt, but the fitness evaluation rewards them for simply making legal moves. Later, when some of the ANNs happen to generate blocking moves, they will be further rewarded. An actual win is rewarded even more. So the two populations co-evolve against easy opponents, like the pre-cheetah and the pre-gazelle.

lle.

Genuine Artificial Intelligence :)

Author: Mitchell Timin
Posted: 08/27/2003