How to train a neural network to supervised data set using pybrain black-box optimization?

Question

I have played around a bit with pybrain and understand how to generate neural networks with custom architectures and train them to supervised data sets using backpropagation algorithm.

However I am confused by the optimization algorithms and the concepts of tasks, learning agents and environments.

For example: How would I implement a neural network such as (1) to classify the XOR dataset using pybrain genetic algorithm (2)?

(1) pybrain.tools.shortcuts.buildNetwork(2, 3, 1)

(2) pybrain.optimization.GA()

Laoighis

You are not the only one who would like to know. – Rectocele

Answer 1

I finally worked it out!! Its always easy once you know how!

Essentially the first arg to the GA is the fitness function (called evaluator in docs) which must take the second argument (an individual, called evaluable in docs) as its only arg.

In this example will train to XOR

from pybrain.datasets.classification import ClassificationDataSet
# below line can be replaced with the algorithm of choice e.g.
# from pybrain.optimization.hillclimber import HillClimber
from pybrain.optimization.populationbased.ga import GA
from pybrain.tools.shortcuts import buildNetwork

# create XOR dataset
d = ClassificationDataSet(2)
d.addSample([0., 0.], [0.])
d.addSample([0., 1.], [1.])
d.addSample([1., 0.], [1.])
d.addSample([1., 1.], [0.])
d.setField('class', [ [0.],[1.],[1.],[0.]])

nn = buildNetwork(2, 3, 1)
# d.evaluateModuleMSE takes nn as its first and only argument
ga = GA(d.evaluateModuleMSE, nn, minimize=True)
for i in range(100):
    nn = ga.learn(0)[0]

Test results after the above script:

In [68]: nn.activate([0,0])
Out[68]: array([-0.07944574])

In [69]: nn.activate([1,0])
Out[69]: array([ 0.97635635])

In [70]: nn.activate([0,1])
Out[70]: array([ 1.0216745])

In [71]: nn.activate([1,1])
Out[71]: array([ 0.03604205])

Laoighis

What does the GA do here? Xor is known to be solvable with simple backprop learning. – Rsfsr

XOR is just a sample dataset. I used it because it is well known, small and simple. GA and BP perform differently depending on the problem. GAs are also nice because they don't have to be used in conjunction with a neural net. – Laoighis