Modifications to BPGEP

I talked to a biology professor today about the biological justification of BPGEP. It turns out that some of the things I thought of to simplify BPGEP are biologically justified. The most interesting is the idea that instead of ensuring that no operation is ever going to create an invalid individual, as PGEP does, or separating each gene into a head and tail, as GEP does, we could just allow any change at any time. Then, when we are doing the evaluation, we repair the gene so that it has a corresponding expression tree. The way I thought of is to have a neutral element that we tac on to fill out the extra needed terminals. Dr. Hibler, my adviser, suggested that instead we just have that any operation that does not have a full complement of terminals just return, unchanged the one(s) we have. This would work for 2-ary operations, but I'm not sure if that would help with n-ary ones (unless we just returned the first argument or something). Apparently the RNA strand that a series of base pairs creates adds a head and a tail called a poly(A) tail, and I hope to add such a tail. In real genetics that tail does not have an effect on the resulting protein (I think) but here it has as little an effect as possible.

The other modification is that apparently sometimes parts of an RNA strand are removed before it creates a protein. This has some effect, but the result is often close to what it would be without the modification. This is apparently not very well understood- but it occurred to me that after expressing parts of the bit vector into symbols, we could remove symbols that encode "bad" things, like the rest of the 2^n symbols that fill out the terminal and operator list in BGEP. This may end up being a bad idea, as we may have individuals that need lots of repairing and symbols that get removed (which allow neutral mutations, so maybe they aren't all bad).

There are a huge number of possible modifications to the GEP system, I'm trying to not get lost in them. Hopefully some of the biologically inspired ones both improve and simplify the system and can be made into some sort of reasonable result. I would rather this not end up being yet another highly complex evolutionary algorithm that no one is interested in because of the high cost of entry (the cost of learning and implementing it).