1
$\begingroup$

I am developing an application to run a genetic algorithm over the input characteristics of a neural network. I am currently looking for help finding a good "genome" to use along with good example problems that might yield interesting results, problems that must have certain types of neural network characteristics to solve.

Currently my genome is pretty sparse and it only consists of altering the networks transfer function along with its hidden layer depth and the widths of each hidden layer. The transfer functions range from

Linear, ramp, step, sigmoid, tanh, guassian, trapezoid, sgn, sin, and log

and the range of the hidden layers count and widths are configurable to whatever is desired.

My current training data is simply the XOR function. I am looking for functions that will test my current genome and suggestions on possible additions to the genome.

$\endgroup$
2
  • 1
    $\begingroup$ This may sound like a wet blanket, but you cannot get more out of machine learning than you put into it. In other words, you cannot design a neural network that works well on any class of problems. If you don't understand this, you will not recognize that your design choices (such as the kind of network transfer functions) will be the significant contributors to over-fitting to training data. So you have to keep in mind that a successful machine learning program is mostly due to the one who designed it for the problems it solves, and hence it is not meaningful to train a neural net to do XOR. $\endgroup$
    – user21820
    Dec 1, 2014 at 6:30
  • $\begingroup$ I think you should also learn about Kolmogorov complexity and how it applies to machine learning and modelling in general, which would help you understand my strange comments. =) $\endgroup$
    – user21820
    Dec 1, 2014 at 6:46

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.