Parameters in Weka Multilayer Perceptron Classifier

Question

I'm doing some experiments with Weka Multilayer Perceptron, and I have some questions relating to its parameters. I've checked the help document but couldn't understand:

• What is nominalToBinaryFilter? How to use?

• normalizeAttribute: I think this is to scale value of features to [-1, 1] range. But how they do it in case the value is not numeric, for example with weather dataset.

• reset: This will reset if the current training process diverges and start again with a lower learning rate. How much should we decrease the current learning rate? (how to identify the next learning rate)

• Initial weights: This isn't a parameter, but how they initialize initial weights? Is it symmetric (something like values inside [-ε, +ε])?

Answer 1

It has been a while since i used WEKA, but here are my comments about bullets 2,3 and 4 which may seem useful to you:

Bullet 2: Normalization is non applicable to categorical (non numerical) attributes so you don't need to worry about this parameter.

Bullet 3: By default reset sets the learning rate to half. Adjustment of learning rate depends on many factors and I suggest searching scholarly articles in case you think you are not covered by the default approach. From my experiecnce,a rule of thumb is to alter learning rate in steps of 0,1

Bullet 4: Initial weights are small random numbers that are not identical

Answer 2

I know this is very old but wanted to add regarding bullet 4: The seed paramter is used to seed a random number generator that is then used to generate the random initial weights. Therefore, if you wanted to explore sensitivity to initial weights, you can use different values here.