Feedforwardnet for XOR problem with logsig transfer function

Hi, I want learn feed forward net for my classification problem. But network doesn't learn anything useful, so I start checking network setting.

I choose xor problem. I find that network doesn't learn xor if the output transfer function is set to logsig! Sample code:

 x = [0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1; 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1];
 y =[0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0];

 net = feedforwardnet([2 2]);
 net = configure(net,x,y);
 net.layers{3}.transferFcn = 'logsig';
 [net,tr] = train(net,x,y);
 net(x)

The outputs are: 0.5000 0.5000 0.9985 0.5000 0.5000 0.5000 0.9985 0.5000 0.5000 0.5000 0.9985 0.5000 0.5000 0.5000 0.9985 0.5000

I try few times and I never got any value smaller then 0.5. If I comment 6th line setting transfer function to logsig the net learn perfect.

Does anyone know why network with logsig doesn't work?

P.S My matlab version MATLAB 7.12.0 (R2011a).

Answer 1

[Greg Heath]

Reduce confusion: Never use feedforwardnet.

Use fitnet for regression or curve-fitting.

Use patternnet for classification or pattern-recognition.

Both call feedforwardnet but have more appropriate defaults and plots.

For quick comparisons of defaults, OMIT semicolons:

 net = fitnet
 net = patternnet
 net = feedforwardnet

For detailed comparisons of source code

 type fitnet
 type patternnet
 type feedforwardnet

----------------------------------------------

 net = feedforwardnet([2 2]);

1. You specified 2 hidden layers instead of 1

 net = configure(net,x,y);

2a. No need to use configure unless you are in a loop. train will automatically initialize if no weights have been assigned.

2b. Stick with MATLAB notation of t for target and y for output

 net.layers{3}.transferFcn = 'logsig';

3. Inappropriate for the default output mapmaxmin to [-1,1]. That is why you are getting the 0.5.

 [net,tr] = train(net,x,y);
  net(x)

4.Use extended LHS and MATLAB t, y notation (semicolons omitted for clarification)

   N            = length(t)
  trueclass     = 1 + round(t)
  net           = patternnet(2)  % One hidden layer
  rng(0)                          % initialize RNG
  [ net tr y ]  = train(net,x,t);
  assignedclass = 1+round(y)
  err           = assignedclass~=trueclass
  Nerr          = sum(err)
  PctErr        = 100*Nerr/N

To mitigate the possibility that rng(0) may not produce a good set of initial weights, use a loop over weight initializations AFTER rng(0) and start the loop using configure. Careful with indexing.

Thank you for formally accepting my answer.

Greg