My current plan for the direction of my thesis is to keep modifying BPGEP until it is as simple and easy to use as possible. The idea is that the original GEP is just the inspiration from which a nice and biologically justified method can be made. There should be no knowledge in the genetic operators about the structure of the tree an individual expresses, nor even about the symbols it encodes. It should be easy to specify a terminal and operator set without getting into details about the binary representation. New genetic operators should be easy to come up with, as all they have to do is preserve length. New chromosome structures (which are common for holding extra information in individuals) should be easy to add- just add to the end like in other EAs. All individuals should encode some expression tree- and if a repairing operation must occur it should be simple and quick. The biological justification should be stressed, as well as the binary representation- PGEP is an amazing method and we have to justify the added complexity of the binary representation if we are going to use it at all (which we are).
I would maybe like some interesting problem to solve that this method is suited for, but if I don't end up with one I expect to perform some symbolic regression task and maybe classification. There are some really cool problems used in the first GEP paper which may be a good testing ground.
It is my hope that in the end the number of parameters that one has to specify is smaller than GEP and that the implementation does not suffer too much complexity.
Looking at the original GEP paper it looks like you need to specify: popsize, Pm, Pc1, Pc2, Pgene recombination, Pris, ris length, Pis, is length, terms, ops, generations, head length, number of genes, linking function, and Pgene transpose. In BPGEP I expect to specify: popsize, individual length, generations, Pm, Pc1, Pc2, Pr, terms, and ops. These are mostly unavoidable and as long as there is a reasonable strategy for dealing with some of the implementation details those may be the only parameters to BPGEP.
It should be noted that a commonly Pc1 = .35, Pc2 = .35, Pr=.6, Pm=.01. The other operators can be set between 50 and 200 somewhere. Terms and ops are problem specific and there is no way around specifying those. This means that there is not all that much to do if we expect the default params to work. I think that this is a desirable property.
ReplyDelete