WinnerTakeAllPopulation.cpp

Go to the documentation of this file.

/************************************************************************
* Verve                                                                 *
* Copyright (C) 2004-2006                                               *
* Tyler Streeter  tylerstreeter@gmail.com                               *
* All rights reserved.                                                  *
* Web: http://verve-agents.sourceforge.net                              *
*                                                                       *
* This library is free software; you can redistribute it and/or         *
* modify it under the terms of EITHER:                                  *
*   (1) The GNU Lesser General Public License as published by the Free  *
*       Software Foundation; either version 2.1 of the License, or (at  *
*       your option) any later version. The text of the GNU Lesser      *
*       General Public License is included with this library in the     *
*       file license-LGPL.txt.                                          *
*   (2) The BSD-style license that is included with this library in     *
*       the file license-BSD.txt.                                       *
*                                                                       *
* This library is distributed in the hope that it will be useful,       *
* but WITHOUT ANY WARRANTY; without even the implied warranty of        *
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the files    *
* license-LGPL.txt and license-BSD.txt for more details.                *
************************************************************************/

#include "WinnerTakeAllPopulation.h"
#include "Neuron.h"

namespace verve
{
        WinnerTakeAllPopulation::WinnerTakeAllPopulation()
        : UltraSparseCodePopulation()
        {
        }

        WinnerTakeAllPopulation::~WinnerTakeAllPopulation()
        {
        }

        void WinnerTakeAllPopulation::updateFiringRatesWTA()
        {
                // IDEA: This could be a little more efficient if we could avoid 
                // updating all of the Neurons' targets (post-synaptic Neurons) 
                // until the end.  Right now, we are updating all firing rates 
                // twice.

                // Compute new firing rates from the synaptic input sums.  Use 
                // a sigmoidal activation.
                unsigned int size = (unsigned int)mNeurons.size();
                for (unsigned int i = 0; i < size; ++i)
                {
                        mNeurons[i]->updateFiringRateSigmoid();
                }

                // Now pick the "winner" Neuron; set its firing rate to 1 and the 
                // rest to 0.  Save the index of the winner.
                mActiveIndex = computeWinnerIndex();

                // Set the winner's firing rate to 1 and the rest to 0.
                for (unsigned int i = 0; i < size; ++i)
                {
                        if (i == mActiveIndex)
                        {
                                mNeurons[i]->setFiringRate(1);
                        }
                        else
                        {
                                mNeurons[i]->setFiringRate(0);
                        }
                }
        }

        unsigned int WinnerTakeAllPopulation::computeWinnerIndex()
        {
                unsigned int winnerIndex = 0;
                unsigned int size = (unsigned int)mNeurons.size();

                // We will use a roulette selection method to choose the winner.  
                // First we find the total firing rate and count the number of 
                // inactive Neurons.
                real totalFiringRate = 0;
                unsigned int numInactiveNeurons = 0;
                for (unsigned int i = 0; i < size; ++i)
                {
                        real firingRate = mNeurons[i]->getFiringRate();
                        if (0 == firingRate)
                        {
                                ++numInactiveNeurons;
                        }
                        else
                        {
                                totalFiringRate += firingRate;
                        }
                }

                // We need to add a little to the total for each inactive Neuron.  
                // This ensures that the 'defaults::minActionSelectionProb' exactly 
                // specifies the minimum action selection probability.
                real extra = 0;
                if (defaults::minActionSelectionProb > 0)
                {
                        // The purpose of using this extra variable is to avoid 
                        // compiler (e.g. gcc 3.4.4) warnings about division by 
                        // zero.  For example, when defaults::minActionSelectionProb 
                        // is zero, the calculation below looks like a divide by 
                        // zero problem, but this will never happen because of 
                        // the 'if' check above.
                        real minProb = defaults::minActionSelectionProb;

                        // The restriction on 'defaults::minActionSelectionProb' ensure 
                        // that the following quantity is a valid positive number.
                        extra = totalFiringRate / (1 / minProb - numInactiveNeurons);
                }
                totalFiringRate += numInactiveNeurons * extra;

                // Generate a random number uniformly between 0 and 
                // 'totalFiringRate.'
                real randomValue = globals::randomRealUniform(0, 
                        totalFiringRate);

                // Pick a "winner" probabilistically.  A higher firing 
                // rate means a higher chance of being chosen as the 
                // winner.
                real currentTotal = 0;
                for (unsigned int i = 0; i < size; ++i)
                {
                        // We need to add some activation to the inactive Neurons to 
                        // make the 'defaults::minActionSelectionProb' valid.
                        real firingRate = mNeurons[i]->getFiringRate();
                        if (0 == firingRate)
                        {
                                firingRate = extra;
                        }

                        if (randomValue <= firingRate + currentTotal)
                        {
                                winnerIndex = i;
                                break;
                        }
                        else
                        {
                                currentTotal += firingRate;
                        }
                }

                return winnerIndex;
        }
}

---