Tuesday, 14 October 2008

Polyworld - Using Evolution to Design AI

I have been watching Polyworld - Using Evolution to Design AI on Google Tech Talks.

I like how there is no fitness function. Rather the system uses pure natural selection. It allows us to set up the environment to shape resultant intelligence, language and knowledge.

I like how energy governs a lot of what is happening. Merely existing takes energy. For creatures to stay around, they must act to find food, mate, avoid predators and so on. Being bigger or having larger brains is also more costly which makes sense. Evolution would only use energy on larger creatures and bigger brain creatures if there is a survival benefit.

I am concerned with the first step fallacy. Polyworld is a fantastic first step but the next step, allowing agents to become even smarter, how about that? I don't like how the focus is on genetic complexity, network complexity, rates of development and so on. I prefer if the focus was on intelligent behaviour. I know that these graphs are important but they just don't have much of an impact on me.

Is there any back propagation going on or do the neural networks only change over time through mutation?

Neo-Darwinism is about networks expanding and then pruning unnecessary connections. Sounds good.

The comment near the end about being able to do this more cleanly is interesting. The kin selection suggestion is also a good one. It resonates with ideas from "The Selfish Gene" by Richard Dawkins. The suggestion requires allowing creatures to be better able to observe each other to select good partners.

How can we get the agents to become incrementally smarter?

Polyworld may be the way to go because of all the work there on the world simulation.

No comments:

Post a Comment