////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/// evaluate
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
RealVector MultiLayerPerceptron::evaluate( const RealVector& x )
{
/// Assert validity of input.
assert( !m_useBiasNodes && x.size() == m_input.size() || m_useBiasNodes && x.size() == m_input.size() - 1 );
/// Set input.
for ( size_t i = 0; i < x.size(); ++i )
{
m_input[ i ] = x[ i ];
}
/// Simple forward propagation when there are no hidden layers.
if ( m_y.size() == 0 )
{
propagateForward( m_input, m_output, m_weights[ 0 ] );
}
else
{
/// Propage from input node to first hidden layer.
propagateForward( m_input, m_y[ 0 ], m_weights[ 0 ] );
applyActivationFunc( m_y[ 0 ], m_x[ 0 ] );
/// Propagate to last hidden layer.
for ( size_t iHiddenLayer = 0; iHiddenLayer < m_y.size() - 1; ++iHiddenLayer )
{
propagateForward( m_x[ iHiddenLayer ], m_y[ iHiddenLayer + 1 ], m_weights[ iHiddenLayer + 1 ] );
applyActivationFunc( m_y[ iHiddenLayer + 1 ], m_x[ iHiddenLayer + 1 ] );
}
/// Propagate to output nodes.
propagateForward( m_x.back(), m_output, m_weights.back() );
}
return m_output;
}
double RecurrentNN::nodeDifference(MNIndividual *ind)
{
RecurrentNN *other = dynamic_cast<RecurrentNN *>(ind);
if (!other)
return INFINITY;
long length = std::max(other->nodes.size(), nodes.size());
double d = 0;
for (long i=0; i<length; i++) {
double d11 = 0;
double d21 = 0;
double d12 = 0;
double d22 = 0;
if (i < nodes.size()) {
d11 = applyActivationFunc(nodes[i], 0) - applyActivationFunc(nodes[i], 1);
d21 = applyActivationFunc(nodes[i], -1) - applyActivationFunc(nodes[i], 0);
}
if (i < other->nodes.size()) {
d12 = applyActivationFunc(other->nodes[i], 0) - applyActivationFunc(other->nodes[i], 1);
d22 = applyActivationFunc(other->nodes[i], -1) - applyActivationFunc(other->nodes[i], 0);
}
d += fabs((d11 - d21)/2 - (d12 - d22)/2);
}
return d;
}
std::vector<double> RecurrentNN::nodeOutputsForInputs(std::vector<double> inputs, std::vector<std::vector<double> > &memory)
{
std::vector<double> newOutputs(nodes.size());
std::deque<long> toVisitNodes;
for (int i=0; i<nodes.size(); i++)
toVisitNodes.push_back(i);
std::set<long> resolvedNodes;
int runs = 0;
while (toVisitNodes.size() > 0) {
long n = toVisitNodes.front();
bool isInput = (inputNodes.find(n) != inputNodes.end());
std::vector<DelayEdge> inputEdges = inputsToNode(n);
double inputSum = nodes[n].bias;
bool tryAgain = false;
for (long j=0; j<inputEdges.size(); j++) {
DelayEdge &e = inputEdges[j];
long from = e.nodeFrom;
long delay = e.delay;
std::vector<double> history = memory[from];
bool isSelfInspection = false;
if (from == n) {
// we're looking at ourselves
assert(delay > 0);
isSelfInspection = true;
}
bool longEnoughHistory = (history.size() > delay) ;
// if (longEnoughHistory) {
// if (std::count(resolvedNodes.begin(), resolvedNodes.end(), from) > 0 || delay > 0) {
// std::vector<double>::iterator valuePos = history.end() - (delay +1);
// assert(!isUnreasonable(*valuePos));
// inputSum += *valuePos;
// } else {
// tryAgain = true;
// break;
// }
// }
if (longEnoughHistory) {
// we've already resolved the output of this node
// the values can be taken from memory
std::vector<double>::iterator valuePos = history.end() - (delay+1);
inputSum += *valuePos;
} else {
if (longEnoughHistory) {
// we should have an answer, but we don't - reschedule for later
tryAgain = true;
break;
}
}
}
if (!tryAgain) {
// all input nodes resolved
resolvedNodes.insert(n);
if (isInput) {
inputSum += (inputs[n]); // the first [inputNodes.size()] nodes are the inputs
}
double output = applyActivationFunc(nodes[n], inputSum);
assert(!isUnreasonable(output));
newOutputs[n] = output;
std::vector<double> &thisMemory = memory[n];
thisMemory.push_back(output);
} else {
// save this for later
toVisitNodes.push_back(n);
}
toVisitNodes.pop_front();
runs++;
}
return newOutputs;
}