Simulation could take into account all of the drawbacks, by making those portions of the simulation noisy, thereby not allowing evolution to depend on those parameters. This idea is call "Minimal Simulations", and I'm sure I've not done the subject any justice whatsoever!!!
A guy called Nick Jakobi was just finishing his thesis when I was studying at Sussex, link to thesis.
Check out the CCNR - Autonomous Robots research page (A graduate research group at Sussex): CCNR
Emergent Behaviour always seemed credible to me (I did my undergraduate degree @ Sussex Uni;), Insect like navigation has been evolved using relatively simple artificial neural networks (you can easily make a simple robot that avoids objects - which could be seen as hate - as Braitenberg said, IIRC).
I was extremenly fortunate enough to be taught by Adrian Thompson (he's an extremely affable guy, btw) at Sussex uni.
IIRC the GA exploited the physical properties of the silicon. What I found interesting was the fact there was a tiny portion of the FPGA, that was not connected to the main 1/3, when removed the chip would not work.
Also, IIRC, a big issue surrounding GA's, is choosing an encoding scheme - deciding gene location and composition.
For my final year proj, I evolved flocking behaviour for Braitenberg type virtual robots, the idea was to evolve an Artificial Neural Network that would control the left and right motors from minimal environmental input (it actually worked - once:)
I also remember a demo that was given showing an eight legged robot that had had ANNs evolved that made it walk, climb over and avoid obstacles. This kind of application of GA's, moves them away from (though not entirely) being a search technique and may perhaps make them useful tools rather than a different way to seach a problem space.
Anyhow, this research is 4 or 5 years old (at least) by now!
I thought GA's might be an approach to making bots for Quake and the like (I always bring it back down to gaming - sorry!), with the fitness of an evolved bot being tested by playing against real people.
if any of you get the chance to go to COGS (the School of Cognitive and Computing Sciences) @ Sussex Uni, I heartily recommend it, fabulous place, and fabulous people!
Nice
Jaymz
Slashdot Mirror
Simulation could take into account all of the drawbacks, by making those portions of the simulation noisy, thereby not allowing evolution to depend on those parameters. This idea is call "Minimal Simulations", and I'm sure I've not done the subject any justice whatsoever!!!
A guy called Nick Jakobi was just finishing his thesis when I was studying at Sussex, link to thesis.
It was an evolutionary robotics thesis.
Check out the CCNR - Autonomous Robots research page (A graduate research group at Sussex):
CCNR
Emergent Behaviour always seemed credible to me (I did my undergraduate degree @ Sussex Uni;), Insect like navigation has been evolved using relatively simple artificial neural networks (you can easily make a simple robot that avoids objects - which could be seen as hate - as Braitenberg said, IIRC).
I was extremenly fortunate enough to be taught by Adrian Thompson (he's an extremely affable guy, btw) at Sussex uni. IIRC the GA exploited the physical properties of the silicon. What I found interesting was the fact there was a tiny portion of the FPGA, that was not connected to the main 1/3, when removed the chip would not work. Also, IIRC, a big issue surrounding GA's, is choosing an encoding scheme - deciding gene location and composition. For my final year proj, I evolved flocking behaviour for Braitenberg type virtual robots, the idea was to evolve an Artificial Neural Network that would control the left and right motors from minimal environmental input (it actually worked - once :)
I also remember a demo that was given showing an eight legged robot that had had ANNs evolved that made it walk, climb over and avoid obstacles. This kind of application of GA's, moves them away from (though not entirely) being a search technique and may perhaps make them useful tools rather than a different way to seach a problem space.
Anyhow, this research is 4 or 5 years old (at least) by now!
I thought GA's might be an approach to making bots for Quake and the like (I always bring it back down to gaming - sorry!), with the fitness of an evolved bot being tested by playing against real people.
if any of you get the chance to go to COGS (the School of Cognitive and Computing Sciences) @ Sussex Uni, I heartily recommend it, fabulous place, and fabulous people!
Nice
Jaymz