My friend Philip and I are still working toward an implementation of a process for evolving artificial neural networks (ANNs), for the domain of game AI (specifically, the Asian board game Go). We worked a bit more on the software that will evolve the neural nets last night.

We're using a variant of an approach developed by researchers at the University of Texas at Austin, Drs. Stanley and Miikkulainen. It's called NEAT (NeuroEvolution of Augmenting Topologies).

If you're curious, and you enjoy reading computer science journal articles, you can find an explanation of NEAT in PDF format here.

The basic idea is starting with small network composed of artificial neurons, or "nodes", connected to each other with various connection weights (emulating biological synapses). NEAT starts out with minimal architectures (few nodes and connections). A population of many networks, e.g. 100, is produced. They compete with each other for a specific task (in our case, they play games against one another). Each network is given a score. As in nature, the best performers survive, and the losers die. The winners produce new offspring by "mating", using crossover and mutation. There are three types of mutation in NEAT: creating a new node, creating a new connection, and changing the value of an existing weight.

Over successive generations, better-performing networks evolve.

Anyway, if you want more detail, check out the paper

We're initially going to apply our methods to simple games (Tic-Tac-Toe, possibly Connect Four, then up to Pente, possibly Hex, and eventually Go). I'll probably blog from time to time on our progress.