nelson_p@apollo.HP.COM (Peter Nelson) (02/08/90)
Earlier I posted: > Are there any general principals that may be applied to > problems like this [microscopic behavior resulting in > macroscopic structure] in simulations? If I wanted to > create a CA program that would produce a particular shape, > say a star or a triangle, is there any systematic way to go > about it? I've been wondering if anyone has taken an evolutionary approach to this? Suppose we populate a CA universe with cells that obey *different* rules? Moreover, suppose we add some source of variablity to the rules, like "mutations" or daughter cells that have some combination of their (not neccesarily just 2) parent's cells. All of the above would probably not be hard to implement. The crucial part, which might also be the hard part, would then be to have some way to "train" this soup. Say we wanted to evolve cells that would form star-shaped structures. If we could find a way to favor those cells that exhibit this characteristic, then over time we could end up with a universe that had cells which mostly formed stars (or whatever). An added problem to the one of providing the "feedback" to our simulation is that of identifying "right" results: Such a system will not go straight from chaos to stars; you will have to "shape" the behavior in small increments, and recognizing the "right" behavior in its earlier forms will be difficult. Finding some way to automate the shaping function would also be helpful since this process could take days or more of computer time at it would be nice to just let it run unattended. ---Peter