Just a quick update on the meta-neat / parameter search work I've been doing. I've had a working version for about a week now and have just started ...
I want to start looking into using a range of different functions with neurons in the hope that this will expand NEAT's ability to find solutions to particular...
... Mean? Median? Mode? ... Some sort of odd-numbered root (cube? fifth?) ... Offered without prejudice, I don't know how many extra dimensions you'd be...
... Mode is an interesting suggestion. I can see that it might be useful, but how do you calculate it when you have a set of floating point numbers? I guess...
I just had another idea on this. During a NEAT search my intention was to not mutate any of the neuron functions once a neuron had been created, primarily...
... Yeh I remember you saying this a while ago but never got around to looking into it. Do you use plain tanh e.g. y=tanh(x) , or some variation? The reason I...
... Quick update. I have just got the average number of gens to find a solution for XOR down to about 40, which is a nice result and also means I won't have to...
... Yes. Rescaling the "vanilla" sigmoid to [-1,1] by multiplying by 2 and subtracting 1 gets you, with a spot of trivial reworking, to (1 exp(-x)) / (1 +...
... Oh? I thought this was just the beginning. At least take your found parameters from XOR and try them on some other problem domains. Ian ... Living@Home -...
Hi, This version looks fairly expensive and I'm not sure exactly what kind of curve it's going to give you. The version I was suggesting is basically an...
... Thanks Mike. With this in mind it seems odd that tanh takes longer to compute than the sigmoid function since the result of exp(-x) can be re-used and they...
Hi, Whilst I'm glad to see this idea getting explored. I wonder whether you need to think through what the objective is. ... (which I guess was originally...
Hi, Anybody know of any work modelling an ANN as a system of PDEs? e.g. dA/dt = k(weighted_sum(inputs) - A) (A = activation before sigmoid) That throws away...
Choosing a good set of parameter is a problem for me also. I would be interested in the details of your method. Have you posted any information about that? I...
All, This is slightly off-topic, but I figure you guys are the best ones to ask on this subject... I am trying to envision a network model where the...
... I'm happy to open out the discussion. The objective is slightly vague but I started out thinking about how we could modify NEAT so that it can solve...
... Sorry, I literally meant finish off for the day :) The same code can also be re-used to run one off benchmarks, e.g. how well does this set of parameters...
... Hi Mitchell, I started out wanting to search for optimal connection weight mutate schemes, and I did post an explanation of how I was going to do that a ...
... OK, I'm with you. ... Do they? There are a few different types of neurones in biology, it is true, but in some ways it is more as if each type plays a...
... <snip> ... Yes, Colin I would appreciate that. I don't want to see code, just an explanation of how it works, with enough detail that I might code...
Hi Colin-- I'm curious if you've looked at the shape of the curves you get from those runs. Do you see nice smooth curves where a small change in a parameter...
... Hi Brian, I don't have any graphs but yes I did notice some smooth curves while adjusting some of the parameters manually and noting numbers on scraps of...
Forgot to say that those parameters I mentioned brought the average for XOR down to 21 generations in the end, which is actually a lower number than that...
... So what you're saying here is that any given circuit is only made up of one type of neurone... ... I think I'm reading you wrong, this seems to contradict...
... No, I think I fumbled that paragraph, disregard it.... ... w.r.t development in natural brains... Development is a whole complex and fascinating subject in...
... What was the text? Any good? ... *Biophysics of Computation: Information Processing in Single Neurons (Computational Neuroscience), Chistoff Koch* *Book...
... Yeh me too. There does seem to be a lot of highly specialised research in the field of biological neurons, but most of it is extremeley hard to digest or...
... For my DelphiNEAT implementation, it takes about 16,45 generations to find a perfect solution (just ran it, averaged over 100 iterations, popsize=150). But...
