source: cpp/gdk/neuroimpl-fuzzy.cpp @ 32

// This file is a part of Framsticks GDK library.
// Copyright (C) 2002-2006  Szymon Ulatowski.  See LICENSE.txt for details.
// Refer to http://www.frams.alife.pl/ for further information.

#include "neuroimpl-fuzzy.h"
#include "neuroimpl-fuzzy-f0.h"

int NI_FuzzyNeuro::countOuts(const Model *m, const Neuro *fuzzy)
{
  int outputs=0;
  for(int i=0;i<m->getNeuroCount();i++)
     for(int in=0;in<m->getNeuro(i)->getInputCount();in++)
        if (m->getNeuro(i)->getInput(in)==fuzzy) outputs++;
  return outputs;
}

int NI_FuzzyNeuro::lateinit()
{
  int i, maxOutputNr;

  //check correctness of given parameters: string must not be null, sets&rules number > 0
  if((fuzzySetsNr<1)||(rulesNr<1)||(fuzzySetString.len()==0)||(fuzzyRulesString.len()==0))
    return 0; //error

  // this part contains transformation of fuzzy sets
  fuzzySets = new double[4*fuzzySetsNr]; //because every fuzzy set consist of 4 numbers
  // converts fuzzy string from f0 to table of fuzzy numbers type 'double'
  // (fill created space with numbers taken from string)
  // also checks whether number of fuzzy sets in the string equals declared in the definition
  if (FuzzyF0String::convertStrToSets(fuzzySetString, fuzzySets, fuzzySetsNr) != 0)
    return 0; //error

  // this part contains transformation of fuzzy rules and defuzzyfication parameters
  rulesDef = new int[2*rulesNr];    //for each rule remembers number of inputs and outputs
  //check correctness of string and fill in the rulesDef
  if (FuzzyF0String::countInputsOutputs(fuzzyRulesString, rulesDef, rulesNr) == 0)
  {
    defuzzParam = new double[rulesNr]; // parameters used in defuzyfication process
    // create space for rules according to rulesDef
    rules = new int*[rulesNr];   //list of rules...
    for (i=0; i<rulesNr; i++)    //...that contains rules body
    {
      rules[i] = new int[2*(rulesDef[2*i]+rulesDef[2*i+1])];  //each rule can have different number of inputs and outputs
      defuzzParam[i] = 0; //should be done a little bit earlier, but why do not use this loop?
    }
    // fill created space with numbers taken from string
    if (FuzzyF0String::convertStrToRules(fuzzyRulesString, rulesDef, rules, fuzzySetsNr, rulesNr, maxOutputNr) != 0)
      return 0; //error
  }
  else
    return 0; //error

  setChannelCount(countOuts(neuro->owner, neuro));
  return 1; //success
}

NI_FuzzyNeuro::~NI_FuzzyNeuro()
{
  if(rules) //delete rows and columns of **rules
  {
    for (int i=0; i<rulesNr; i++) SAFEDELETEARRAY(rules[i])
    SAFEDELETEARRAY(rules)
  }
  SAFEDELETEARRAY(defuzzParam)
  SAFEDELETEARRAY(rulesDef)
  SAFEDELETEARRAY(fuzzySets)
}

int NI_FuzzyNeuro::GetFuzzySetParam(int set_nr, double &left, double &midleft, double &midright, double &right)
{
  if ( (set_nr>=0) && (set_nr<fuzzySetsNr) )
  {
    left = fuzzySets[4*set_nr];
    midleft = fuzzySets[4*set_nr+1];
    midright = fuzzySets[4*set_nr+2];
    right = fuzzySets[4*set_nr+3];
    return 0;
  }
  else
    return 1;
};

/** Function conduct fuzzyfication of inputs and calculates - according to rules - crisp multi-channel output */
void NI_FuzzyNeuro::go()
{
  if (Fuzzyfication()!=0)
    return;
  if (Defuzzyfication()!=0)
    return;
};

/**
* Function conduct fuzzyfication process - calculates minimum membership function (of every input) for each rule,
* and writes results into defuzzParam - variable that contains data necessary for defuzzyfication
*/
int NI_FuzzyNeuro::Fuzzyfication()
{
  int i, j, nrIn, inputNr, nrFuzzySet;
  double minimumCut; // actual minimal level of cut (= min. membership function)

  // sets defuzzyfication parameters for each rule:
  for (i=0; i<rulesNr; i++)
  {
    nrIn = rulesDef[2*i]; // nr of inputs in rule #i
    minimumCut = 2; // the highest value of membership function is 1.0, so this value will definitely change
    for (j=0; (j<nrIn)&&(minimumCut>0); j++) //minimumCut can not be <0, so if =0 then stop calculations
    {
      nrFuzzySet = rules[i][j*2 + 1]; // j*2 moves pointer through each output, +1 moves to nr of fuzzy set
      inputNr = rules[i][j*2]; // as above but gives input number
      minimumCut = min( minimumCut, TrapeziumFuzz(nrFuzzySet, getWeightedInputState(inputNr))); // value of membership function for this input and given fuzzy set
    }
    if ( (minimumCut>1) || (minimumCut<0) )
      return 1;
    defuzzParam[i] = minimumCut;
  }
  return 0;
};

/**
* Function calculates value of the membership function of the set given by wchich_fuzzy_set for given crisp value input_val
* In other words, this function fuzzyficates given crisp value with given fuzzy set, returning it's membership function
* @param which_fuzzy_set - 0 < number of set < fuzzySetsNr
* @param input_val - crisp value of input in range <-1; 1>
* @return value of membership function (of given input for given set) in range <0;1> or, if error occur, negative value
*/
double NI_FuzzyNeuro::TrapeziumFuzz(int which_fuzzy_set, double input_val)
{
  double range=0, left=0, midleft=0, midright=0, right=0;

  if ( (which_fuzzy_set < 0) || (which_fuzzy_set > fuzzySetsNr) )
    return -2;
  if ( (input_val < -1) || (input_val > 1) )
    return -3;

  if (GetFuzzySetParam(which_fuzzy_set, left, midleft, midright, right) != 0)
    return -4;

  if ( (input_val < left) || (input_val > right) ) // greather than right value
    return 0;
  else if ( (input_val >= midleft) && (input_val <= midright) ) // in the core of fuzzy set
    return 1;
  else if ( (input_val >= left) && (input_val < midleft) ) // at the left side of trapezium
  {
    range = fabs(midleft - left);
    return fabs(input_val-left)/((range>0)?range:1); // quotient of distance between input and extreme left point of trapezium and range of rising side, or 1
  }
  else if ( (input_val > midright) && (input_val <= right) ) // at the right side of trapezium
  {
    range = fabs(right - midright);
    return fabs(right-input_val)/((range>0)?range:1); // quotient of distance between input and extreme right point of trapezium and range of falling side, or 1
  };

  // should not occur
  return 0;

};

/**
* Function conducts defuzzyfication process: multi-channel output values are calculates with singleton method (method of high).
* For each rules, all outputs fuzzy sets are taken and cut at 'cut-level', that is at minumum membership function (of current rule).
* For all neuro pseudo-outputs, answer is calculated according to prior computations.
* In fact, there is one output with multi-channel answer and appropriate values are given to right channels.
*/
int NI_FuzzyNeuro::Defuzzyfication()
{
  int i, j, nrIn, nrOut, out, set, outputsNr;
  double *numerators, *denominators, midleft, midright, unimp;

  outputsNr = getChannelCount();

  numerators = new double[outputsNr];
  denominators = new double[outputsNr];

  for(i=0;i<outputsNr;i++) numerators[i] = denominators[i] = 0;

  // for each rule...
  for (i=0; i<rulesNr; i++)
  {
    nrIn = rulesDef[2*i]; // number of inputs in rule #i
    nrOut = rulesDef[2*i + 1]; // number of outputs in rule #i
    // ...calculate each output's product of middle fuzzy set value and minimum membership function (numerator) and sum of minimum membership function (denominator)
    for (j=0; j<nrOut; j++)
    {
      out = rules[i][2*nrIn + 2*j]; //number of j-output
      set = rules[i][2*nrIn + 2*j + 1]; //number of fuzzy set attributed to j-output
      if (GetFuzzySetParam(set, unimp, midleft, midright, unimp) != 0) // gets range of core of given fuzzy set
        { SAFEDELETEARRAY(denominators) SAFEDELETEARRAY(numerators) return 1; }
      //defuzzParam[i] = minimum membership function for rule #i - calculated in fuzzyfication block
      // defuzzyfication method of singletons (high): (fuzzy set modal value) * (minimum membership value)
      numerators[out] += ((midleft + midright)/2.0) * defuzzParam[i];
      denominators[out] += defuzzParam[i];
    }
  }

  for (i=0; i<outputsNr; i++)
  {
    if (denominators[i] == 0)
      setState(0, i);
    else
      setState(numerators[i]/denominators[i], i);
  }

  SAFEDELETEARRAY(denominators)
  SAFEDELETEARRAY(numerators)

  return 0;
};