Dave Harris wrote: > Mitchell Timin wrote: > >> Dave Harris wrote: >> >>> I began work this morning on wbreduce (waybak reduce), similar to the >>> 4Play >>> examine program. The program tries to eliminate neurons, and failing >>> that, >>> attempts to eliminate one or more weights. This program only applies to >>> perfect ANNs. >>> >>> There are some differences (for example, 20 sets of weights in the >>> chromosome file, of which one (or more) is a perfect ANN.) I haven't >>> encountered anything I don't understand yet. I should have something >>> operational today or tomorrow. >>> >>> >> That's great, Dave. I'm really glad you are working on that. About the >> 20 chromes per file, the examine program just reads the first chrome, >> and that is the best one. I'm pretty sure that's the way the 4Play code >> is, and it should be the same for waybak. >> >> In 4play, we had hundreds of weights, and many could be eliminated. And >> the same for the dozens of neurons. >> Here, with only 4 neurons, I don't think it's possible to eliminate one, >> but there's a good chance to eliminate a few weights. >> >> > I have the preliminaries working, but I got a surprise. When I evaluate the > ANNs in 3-3s-4n-perf.sc, all 20 of them yield a perfect fitness rating! I've > dumped the weights and I can see that all of them are unique. If its true, > then I need to test all ANNs for reducibility, not just the first one in the > file. > I would expect that they differ only slightly, because I think the best ANN would have reproduced itself until it filled the population, except for mutations that occur in the final generation. But it's also possible to have multiple ANNs of equal fitness whose chromosomes seem very different. They may not really be different if the neurons and weights were sorted in some way. For example, given an ANN and its chromosome, you could swap two neurons and all of their connections, and then the chromosome would seem quite different, the the two ANNs are equivalent. (Imagine the ANNs were made of hardware, with long loose wires. You could rearrange the neurons on your table, and the wire connection would not change. But the chromosome describing the ANN would change.) > Time for a reality check - is my code working? I fed it 3-3s-4n-8743.sc and > it rejected every ANN (as expected.) There could be an infinitude of 4 > neuron ANNs that solve the 3-3s problem. > That is correct. In general, there are very many different ANNs that will produce a given fitness rating, even a maximal rating. > I'm proceeding to the reduction code now. > Excellent!
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
