nataraj chakraborty wrote: > Yes I am... > I suppose we should not totally give up our approaches since I have > already figured out the fourmule for type of SCRA, it's best to keep > the design of every project such that either ANNEvolve or NEAT can be > plugged in when ever required. > Problem with me is that I don't ususlly find as much free time as when > I joined ANNEvolve, as such I most of the times slower than required. > Mitchel I have few doubts regarding 4Play changing to SCRA type: > > 1 > Tell which are the files I need to change most. > My guess is Neuron.h/c and Evolve.h/c > > 2 > The choromosome structure of SCRA which I would like to use and > which I have arrived after a lot of thought is different fron what we had. > it will be 2-d array Jagged array kind of thing, I mean having 2 Rows, > the first row carrying carrying the respective weights and each of > this will have an associated set which will sotre fan out sets. > I am still not very much comfortable of having an fan-out sets but > there seems to be no other idea to save calculation. > > 3 > Tell me when and where you actually introduce new Networks into > the evolution, since during that period of time we have to calculate > some equations to know whether the network will be FCBA or Pruned or > Perfect. > Formulas are something like this : > > mN / (s + 1) where N is total number of node > s is skip > m is integer, 1,2,3.... > find the minimum m for which the above fraction gets prefectly divisible. > > e.g, for N = 27 and S = 11 we have m * 27 / (11 + 1) = m27 / 12 = m 9 / 4 > => m = 4 > Now let FO = (float) N / (s +1 ) = 27 / 12 = 2.25 > > Now is FO * M < N - 1 then this is not FCBA > or else this is FCBA > thus 2.25 * m = 2.25 * 4 = 9 << 27 => Not FCBA > > Next it remain to determine whether this is a pruned network > > the formule is > 1 + (s + 1)n = (x - 1)A + x > where A is advanced... > and n and x are integers and independent variable. > if for and pair of (n,x) the above equality holds than the above > network is pruned > or else it is perfect and desiarable. > > I have verified the above conjecture with networks of size 30 and is > confident that above law holds although I have no methatical prove for it. > > 4 > Now the question of crossover, If the chromosome if as mentioned > in point 2 than it is all about exchanging splice. > Since if we cross two networks having follwoing parameters > (27,11,1) and (32,9,3) > that the essential length of both can something like > x+ a and y + b > after crossover > x+b and y + a such that the length remains the same, we'll not > exchange the fab-out sets for the above but only th weights. > Further correction of this idea is needed. > 5 > when and where mutation occurs and also what are the thinga to > mutate on, I mean N or S or A... I feel only N can be mutated. > > I need some help redarding these from you. > work out the above formulas if possible, I will let you know further > developmenets from my part. > I would also like to know if i would like to write the 4Play in a such > a manner so that with a change of parameter it can become 8 play or 12 > play what needs to be done... If the chromosome length is constant, then only neuron.c/h need to be changed. There is also the formula for the length of the chromosome; that needs to be changed. I don't remember which file it's in. You will prolly need parameters for the sparseness.
I suggest that the first goal is to test it with constant chromosome length. This means that the topology would be fixed, but it would be a sparser net than currently used in 4Play. If the topology can vary during evolution, then the chromosome length changes, and this requires somewhat complex additions to evolve.c/h. m -- I'm proud of http://ANNEvolve.sourceforge.net. If you want to write software, or articles, or do testing or research for ANNEvolve, let me know. ------------------------------------------------------------------------- Using Tomcat but need to do more? Need to support web services, security? Get stuff done quickly with pre-integrated technology to make your job easier Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642
