%NEURC Automatic neural network classifier% % 	W = NEURC (A,UNITS)%% INPUT%   A     Dataset%   UNITS Array indicating number of units in each hidden layer (default: [5])%% OUTPUT%   W     Trained feed-forward neural network mapping%% DESCRIPTION% Automatically trained feed-forward neural network classifier with% lenght(UNITS) hidden layers of UNITS(I) units in layer I. Training, by LMNC,% is stopped when the performance on an artificially generated tuning set of % 1000 samples per class (based on k-nearest neighbour interpolation) does not % improve anymore. In future versions UNITS might be optimized automatically% as well. For the time being it is set to 0.2 time the size of the smallest% class in A.%% NEURC always tries three random initialisations, with fixed random seed, % and returns the best result according to the tuning set. This is done in % order to obtain a reproducable result.%% Uses the Mathworks' neural network toolbox.% % SEE ALSO% MAPPINGS, DATASETS, LMNC, BPXNC, GENDATK% Copyright: R.P.W. Duin, duin@ph.tn.tudelft.nl% Faculty of Applied Physics, Delft University of Technology% P.O. Box 5046, 2600 GA Delft, The Netherlands% $Id: neurc.m,v 1.10 2003/12/28 17:32:21 bob Exp $function argout = neurc (a,argin)	prtrace(mfilename);	mapname = 'Automatic Neural Classifier';	n_attempts = 3;							% Try three different random initialisations.	if (nargin < 2)    prwarning(1,'no network architecture specified, assuming one hidden');		argin = []; 	end	if (nargin < 1) | (isempty(a))		w = mapping(mfilename,argin);		argout = setname(w,mapname);		return	end	[m,k] = size(a);	if isempty(argin)		cs = classsizes(a);		argin = ceil(0.2*min(cs));	end		if (~ismapping(argin)) 				islabtype(a,'crisp');		isvaldset(a,1,2); % at least 1 object per class, 2 classes		% Second parameter is not a mapping: train a network.		units = argin;		% Reproducability: always use same seeds. 		rand('seed',1); randn('seed',1); opt_err = inf; opt_mapping = [];		% Try a number of random initialisations.		for attempt = 1:n_attempts			prwarning(4,'training with initialisation %d of %d',attempt,n_attempts);			t = gendatk(a,1000,2,1); 			% Create tuning set based on training set.			w = lmnc(a,units,inf,[],t);		% Find LMNC mapping.			e = t*w*testc;								% Calculate classification error.			if (e < opt_err)							 				% If this is the best of the three repetitions, store it.				opt_mapping = w; opt_err = e;				end		end		% Output is best network found.		argout = setname(opt_mapping,mapname);	else 		% Second parameter is a mapping: execute.		w = argin;															data = getdata(w); 		if (length(data) > 1)	    % "Old" neural network - network is second parameter: unpack.  	  data = getdata(w); weights = data{1};    	pars = data{2}; numlayers = length(pars);	    output = a;                       % Output of first layer: dataset.  	  for j = 1:numlayers-1    	  % Number of inputs (n_in) and outputs (n_out) of neurons in layer J.      	n_in = pars(j); n_out = pars(j+1);	      % Calculate output of layer J+1. Note that WEIGHTS contains both  	    % weights (multiplied by previous layer's OUTPUT) and biases    	  % (multiplied by ONES).      	this_weights = reshape(weights(1:(n_in+1)*n_out),n_in+1,n_out);	      output = sigm([output,ones(m,1)]*this_weights);  	    % Remove weights of this layer.    	  weights(1:(n_in+1)*n_out) = [];	    end		else			% "New" neural network: unpack and simulate using the toolbox.			net = data{1};			output = sim(net,+a')';		end;		% 2-class case, therefore 1 output: 2nd output is 1-1st output.		if (size(output,2) == 1)			output = [output (1-output)]; 		end		% Output is mapped dataset.		argout = setdat(a,output,w);		endreturn