June 16, 2008 17:46:43
Posted By Peter Bentley
Here's the second half of the same interview.
- Regarding robots; are genetic algorithms the best approach to make it move? That is, do they yield the best performance, and aren't they limited by lack of processing power or a long time needed to evolve a movement behaviour?
GAs are a great idea if you want to incorporate ideas of embodiment. In other words, if you want your robot to be able to affect its environment in as many ways as possible, and if you want the environment to affect the robot (resulting in improved body and brain) as much as possible. This is how natural organisms are - they shape their world, and their world shapes them. Evolution enables us to test robots in the real world and has a wonderful ability to exploit everything possible to improve those robots. The downside is of course that we can't really evolve robots. We don't have robots that can have children (or that can build themselves), so if we want to use GAs right now, we have to use a combination of computer simulation and physical testing, which can be slow.
- As a sort of subquestion to the one above, do you think evolutionary algorithms are the way to go to make robots robust for hardware failure?
I think evolution is half of the solution. The other half is development (or embroygenesis). If we evolve a growth process, which generates our desired hardware, then that hardware "knows" what it wants to be. So if it gets damaged, the growth process automatically replaces the damaged elements. This is an important trick that we've only just begun to explore, but we're all very excited about the possibilities.