Mitchell Timin wrote:
We've been using 19 neurons for a while now. The best fitness has been only .83. Meanwhile, after a few overnight runs of metavolv.py, it keeps coming up with much few neurons. I now think that 9, 10 or 11 would be best, although I'm not sure if one of those can reach 100%. I think it's important to reach 100% to prove that this type of ANN can fully handle this fairly simple problem. I also plan to gradually reduce the path length until I get there. If I have to go down to a 3 step path to reach 100%, so be it. For now though, I'm gonna just change from 8 steps to 7 for the upper limit.There is good news and bad news. The good news is that I have an ANN that succeeds 100%. The bad news is that it's with a path length of only 3 steps. This one has six neurons. Working with short paths and small neuron counts is a lot nicer because everything moves very quickly. Runmany.bat only takes a few minutes. Since I'm mainly interested in ANNs that do 100%, I think it's best that we work our way upward from the bottom. Hence I would like to find the ANN with the fewest neurons that can do 100% on the 3-step path.Soon I will have recommendations for how you can help; i.e., how many neurons and what path lengths, what parameter values, etc.If anyone wants to run the meta-evolver for this problem, fine. I will help if requested.
The ANN that I mentioned, which does 100% for 3 steps, was evolved using a range of 3 to 4 steps. I've attached a copy of the savedchromes file which I think is the one I'm talking about, although I'm not certain.
The ANN hits about 94% over all 3 & 4 step paths.There is a reason, now, for the search for a tiny ANN to do 100% on 3 steps. It will make a nice addition to VizANN.py. Then everyone can see in step-by-step detail exactly how this binary ANN functions as a state machine. Dave has expressed an interest in creating this. The problem is, even a six neuron ANN of this type has an awful lot of connections to illustrate, so if we can find one with few neurons, it will be much more practical. I you would like to help in this search, please do. At the moment I won't tell you exactly how to proceed, except that in your waybak.ini file, neurons should be 6 or less, minSteps should be 3, and maxSteps could be 3, 4 or 5. Other than that, I'm not sure what would be most fruitful. My advice is, play around with the problem.
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.
3-4s-6n-9539d.sc
Description: Binary data
------------------------------------------------------------------------- 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
