I revisited my Non-Markov Tic Tac Toe experiment (see previous post), but this time using what I thought to be an impossibly simple architecture; only 2 nodes in the hidden layer (note: I am using Sinusoidal rather than Sigmoid activation functions). Amazingly, I observed it reaching a 43% level of fitness (64% was the highest I had observed before although I have been unable to find that level of fitness on subsequent runs), and it was able to find this solution rather rapidly compared to previous tests.
This astonished me because I would not have thought that such complex behavior could be generated by such a small set of parameters. This implies to me that the billions of neurons in the human brain are not really necessary for creating complex behaviors; I suspect that instead they are primarily used for statistical feature detection and also for memory formation and retrieval. This may make me sound a bit like Wolfram, although I think he is still oversimplifying human intelligence.
This also suggests that humans do not possess much in the way of intuitions regarding algorithms created by the process of evolution (again, hints of Wolfram). This sentinment is echoed by Danny Hillis in his latest book, where he talks about the sorting algorithms that his Connection-Machine discovered, and his inability to understand them analytically. To me this is a hopeful state of affairs, because it may mean that many things that we currently consider to be 'hard' problems may not be hard at all, and may in fact possess solutions that are rather easy for evolutionary methods to discover.