> > From: Mitchell Timin <[EMAIL PROTECTED]> > Date: 2006/08/22 ti AM 12:10:14 CEST > To: Evolution of Artificial Neural Networks > <[email protected]> > Ämne: Re: [annevolve] High neuron-count network > > [EMAIL PROTECTED] wrote: > >>> Hi! > >>> I'm struggling to get an network with many neurons to learn a simple xor. > >>> So far ~30 neurons is ok, but if I double that I get an network that is > >>> more like an random number generator. > >>> > >>> The network characteristic is an fully connected network with a > >>> nonlinear output function. > >>> > >>> Do you have any ideas how to get an stable network that can solve xor > >>> and also have many number > >>> of neurons ? > >>> > >>> > >> XOR only requires 2 neurons if there is feedback, 3 otherwise. If you > >> have lots of extra neurons the network is capable of much more complex > >> behaviour. My suggestion is to increase the complexity of your fitness > >> function so that the network evolves toward the behaviour you want. > >> > >> But why do you want 30 neurons when only 3 are required for the functions? > >> > >> Are you using our XOR software from the releases, or something else? > >> > > > > I'm after solving a more complex and diffuse problem. But before I get > > there I must have some tests to prove that an large ANN can solve a simple > > problem like xor. > > > > I have two problems with annevolve-xor: > > - annevolve's xor can only handle fixed neuron count of 2. > > - I'm missing an simple interface that just prints the xor-table > > using the best ANN in the population, and doing this for every nth epoch. > > > > If you agree on these problems, I could poke with the xor code. > > > > Anyway the original question is still more interesting, and so far I'm > > thinking of doing an test-run foreach ANN with fixed or random input during > > the population initialization phase, and those ANN that isn't stable get > > replaced. > > > Since a small ANN can solve XOR, a large one certainly can, because it > might have mostly dormant neurons, and is therefore equivalent to the > small ANN. You can get a dormant neuron if all of its weights and its > bias are very close to zero. Or you can have an ANN that is equivalent > to multiple copies of the small XOR ANN, in parallel. > > Don't use the ANNEvolve XOR. > > Decide what type of ANN you need. What kind of signals are the inputs? > What kind of signals should the outputs be? > Will feed forward work, or is an internal state required? > > m >
Hi thanks for the thoughts. I'll test the close to zero weights approach. That also means you must use that method for every problem. I want to apply this network to recognize sound waves. The input is a of a stream-type (real-time or recorded, but feeded in blocks). Output is an single neuron that says "pattern recognized" or "pattern not recognized" I'll leave everything up to the network, a fully-connected ann that is only changed by random weight changes. ------------------------------------------------------------------------- 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
