ANNEvolve :: Evolution of Artificial Neural Networks

Most of the history of AI, in my opinion, has been an effort to make computers pretend to be intelligent, to fool people. The great Alan Turing, with his Turing Test proposal, probably is largely responsible for this. He advised that if you could get a computer to fool people into thinking it was a person, then we should call that computer intelligent. I for one don't think too much of this definition. For one thing, people are easily fooled. However, his definition was adopted by the young AI community.

There is another definition of intelligence which I prefer. I'm not sure who first proposed this. Intelligence is when you can figure out something on your own. If you present a human with a new problem, and assuming he is motivated, he will think about it, and try out some different approaches that might work, and perhaps arrive at a solution. So the definition I prefer for intelligence is that the entity involved, human or computer, is able to arrive at a solution without it being taught or demonstrated.

Simulated genetic evolution offers one way to achieve this. The evolution might be applied to algorithms, as in Koza's genetic programming (GP), or it might be applied to ANN's. In either case, if you can get a computer to find a solution without guiding it, but just by waiting until it tries several million times, but being guided by the somewhat mysterious process of evolution, then you have genuine AI.

The genetic algorithm approach, GA, does not usually meet this definition because normally the general form of the solution is established by the programmers, GA being used to determine various parameter values. A GA approach might qualify if the form of the solution was very general, rather than being adapted to a specific problem. That is what we are doing with annevolve. We are using a form of GA with a general rather than a specific form of solution.

GP may or may not meet the definition because it requires the programmers to choose a set of primitives. If this set is very general then I would say the GP qualifies. If the set is carefully adapted to the specific problem at hand then it doesn't.

So if you want the closest thing we have to genuine AI, then work at the automated evolution of neural nets or programs