Evolution Environment
Author
I have been thinking about how to create complex intelligence like in humans. I have come to the conclusion that it would take a very long time to program AI logic that is anywhere near as complex and advanced as that of a Human. This is the ultimate goal of AI. It could also be very hard for us as our own understanding of Intelligence and brain function may be limited more than we know. Neural Networks and Fuzzy Logic could simply be false analogies and not true replications of intelligence.
Wouldn't it be easier and more feasable to create an environment for AI evolution rather than to code the AI by hand? The AI could somehow evolve by itself from nothing just like humans did. There would be no defined 'reward' or feedback, except the fact that if the creatures self-multiplied, they would be able to spread and pass their self-replicating code on to others. Combined with random morphism and parent-child code merging, Natural selection would occur and cause the creatures to become more intelligent and capable.
I have created a few simple codes for testing this and I think this evolution idea seems like it could really work and create AI due to survival of the fittest! Are there any problems with this theory? It may take a while for anything to randomly evolve at first but once simple survival behaviors had been evolved, i think not much could stop the creatures from becoming increasingly more intelligent and capable over time.
I agree with the idea that machine evolution is going to be the path to true machine intelligence, however you're possibly underestimating the strength of the words 'somehow' and 'it may take a while'... :)
I actually had plans, in my youth, before I became old and cynical, to implement something like what you're describing. It's still a fascinating topic... linkies (which you may or may not already have):
Framsticks - 3D evolved creatures. Most of them are silly but some are really cool.
Karl Sims' Homepage - Karl Sims is über.
There was also a group that were using skeleto-muscular systems controlled by neural nets to generate realistic 'virtual stunt men'. I can't find a link for the company themselves but there was a New Scientist article on them, and a siggraph paper by some of the people there. Their demo reel was very impressive - I think I've still got it at work somewhere.
I actually had plans, in my youth, before I became old and cynical, to implement something like what you're describing. It's still a fascinating topic... linkies (which you may or may not already have):
Framsticks - 3D evolved creatures. Most of them are silly but some are really cool.
Karl Sims' Homepage - Karl Sims is über.
There was also a group that were using skeleto-muscular systems controlled by neural nets to generate realistic 'virtual stunt men'. I can't find a link for the company themselves but there was a New Scientist article on them, and a siggraph paper by some of the people there. Their demo reel was very impressive - I think I've still got it at work somewhere.
A major problem is the lack of pressure for the creatures to be intelligent. We're intelligent because being so gave us a survival advantage. We had to compete with other species for the resources we need to survive. We also had to overcome the environment in a sense - things like lack of food, disease, weather, and so on. If your simulation just consists of a group of agents with no pressure, you'll just be spinning wheels.
This is the kind of stuff that I have been working on for about a year now. I am not, however, trying to get anything close to the inteligence of a human. Except for maybe in a very specific problem. Ihave some information about the system I am working on at my website dwiel.no-ip.com, but not much more than the documentation which is built every night.
I tried to do this a couple of years ago.
Although I included a fair number of parameters for the agents, gave them several actions and placed them in an environment with pressure to survive the results as to 'intelligence' are not that rewarding. You do get to see basic principles like Darwinism and survival of the fittest confirmed.
The problem is that the scope is too limited. In order to create true 'intelligence', you have to make your agents almost free to do as they please in an environment where everything is possible (limited by physical constraints of course). But that is quite hard to program because if your agents should be able to do anything, on what programmatorical basis can you examine them to see whether they get more intelligent?
Anyway, you can view the results of the experiment on
[link]http://www.softline.be/StrategicAlliance/ALife/en/index.htm[/link]
Tell me what you think...
Although I included a fair number of parameters for the agents, gave them several actions and placed them in an environment with pressure to survive the results as to 'intelligence' are not that rewarding. You do get to see basic principles like Darwinism and survival of the fittest confirmed.
The problem is that the scope is too limited. In order to create true 'intelligence', you have to make your agents almost free to do as they please in an environment where everything is possible (limited by physical constraints of course). But that is quite hard to program because if your agents should be able to do anything, on what programmatorical basis can you examine them to see whether they get more intelligent?
Anyway, you can view the results of the experiment on
[link]http://www.softline.be/StrategicAlliance/ALife/en/index.htm[/link]
Tell me what you think...
i do agree with a previous post ...
even if you can program a 'matrix' where everything is possible and limits are only physical rules (this is by itself a fantastic goal) you'll have to deal with probabilities and explosion of possibilities (exponential increase at least).
the 'intelligence' of humanity appeared because an incredible amount of variables were at a specific value and is everything but the logic result of a very long determinist equations system. Given the time you'll need to evaluate a ecosystem (the different pressures for example), you'll be very lucky if you can find something 'intelligent' (like a worm).
even if you can program a 'matrix' where everything is possible and limits are only physical rules (this is by itself a fantastic goal) you'll have to deal with probabilities and explosion of possibilities (exponential increase at least).
the 'intelligence' of humanity appeared because an incredible amount of variables were at a specific value and is everything but the logic result of a very long determinist equations system. Given the time you'll need to evaluate a ecosystem (the different pressures for example), you'll be very lucky if you can find something 'intelligent' (like a worm).
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
