bayesian.cc

贝叶斯优化算法是一种新的演化算法

// ################################################################################
//
// name:          bayesian.cc      
//
// author:        Martin Pelikan
//
// purpose:       functions for construction and use of bayesian networks 
//                (not including the metric related functions and some of more
//                specific functions defined elsewhere)
//
// last modified: February 1999
//
// ################################################################################

#include <string.h>

#include "bayesian.h"
#include "boa.h"
#include "population.h"
#include "graph.h"
#include "memalloc.h"
#include "recomputeGains.h"
#include "checkCycles.h"
#include "utils.h"
#include "computeCounts.h"
#include "binary.h"
#include "random.h"

// ================================================================================
//
// name:          constructTheNetwork
//
// function:      constructs the network for a given population of strings (data
//                set), and stores it into the graph structure, satisfying the
//                maximal number of incoming edges constraint defined in the
//                parameters passed to the BOA; uses a simple greedy algorithm
//                with only edge additions allowed adding edges which increase
//                the network most (positively) and do not violate constraints;
//                starts with an empty network
//
// parameters:    population...the population of strings to model (a data set)
//                G............where the resulting network is to be stored
//                boaParams....the parameters sent to the BOA
//
// returns:       (int) 0
//
// ================================================================================

int constructTheNetwork(Population *population, AcyclicOrientedGraph *G, BoaParams *boaParams)
{
  int   k,l;

  int   maxNumberOfEdges,numberOfNodes,maxNumberOfIncomingEdges;

  int   numAdded;
  char  finito;

  float **gain;
  char  *full;

  int   maxFrom, maxTo;
  float maxGain;
  
  // empty the graph

  G->removeAllEdges();
 
  // if no incoming edges allowed, the empty network is the solution

  if (boaParams->maxIncoming==0)
    return 0;

  // set some variables

  numberOfNodes            = G->size();
  maxNumberOfIncomingEdges = boaParams->maxIncoming;

  // maximal number of edges is determined by the number of nodes and the maximal number of incoming edges
  // into each node

  maxNumberOfEdges = numberOfNodes*maxNumberOfIncomingEdges;

  // allocate the array for the gains with respect to the addition of each of the edges

  gain = (float**) Calloc(numberOfNodes,sizeof(float*));
  for (k=0; k<numberOfNodes; k++)
    gain[k] = (float*) Calloc(numberOfNodes,sizeof(float));

  // allocate memory for the array containing information about the full nodes (no more edge can go in those)

  full = (char*) Malloc(numberOfNodes);

  // a greedy algorithm for the network constrution -----------------------------------------------
  
  finito = 0;

  // recompute the gains for edges into all nodes and set each node as not full yet
  
  for (k=0; k<numberOfNodes; k++)
    {
      full[k] = 0;
      recomputeGains(k,gain,full,G,population);
    }

  for (numAdded=0; (numAdded<maxNumberOfEdges)&&(!finito); numAdded++)
    {
      maxGain = -1;
      maxFrom = 0;
      maxTo   = 0;
 
      // find the best edge addition (from maxFrom to maxTo with the increase of a metric maxGain)
     
      for (k=0; k<numberOfNodes; k++)
	for (l=0; l<numberOfNodes; l++)
	  if (k!=l)
	    {
	      // what about an edge from k to l?
	      
	      if (gain[k][l]>maxGain)
		{
		  maxFrom = k;
		  maxTo   = l;
		  maxGain = gain[k][l];
		};
	    }
     
      // add an edge with the best gain over all edge additions if its gain is positive, in the other case
      // the greedy algorithm stops
      
      if (maxGain>0)
	{
	  // add the new edge
	  
	  G->addEdge(maxFrom,maxTo);
	  
	  // the edge can't be added anymore
	  
	  gain[maxFrom][maxTo] = gain[maxTo][maxFrom] = -1;
	  
	  // edges that would create cycles with the new edge have to be disabled
	  
	  checkCycles(maxFrom,maxTo,gain,G);
	  
	  // is the ending node of the new edge full yet?
	  
	  full[maxTo]=(G->getNumIn(maxTo)==boaParams->maxIncoming);
	  
	  // recompute the gains that are needed
	  
	  recomputeGains(maxTo,gain,full,G,population);
	}
      else
	finito = 1;
    };
  
  // ----------------------------------------------------------------------------------------------

  // free the memory used by the gain array

  for (k=0; k<numberOfNodes; k++)
    Free(gain[k]);
  Free(gain);

  // free memory used by an auxilary array "full"

  Free(full);

  // get back

  return 0;
};

// ================================================================================
//
// name:          generateNewInstances
//
// function:      generates new instances of the variables (strings) according to 
//                the joint distribution encoded by the constructed network and the
//                conditional probabilities in a modeled data set
//
// parameters:    parents......the modeled population
//                offspring....where the new instances should be stored
//                G............the network used as a model
//                boaParams....the parameters sent to the BOA
//
// returns:       (int) 0
//
// ================================================================================

int generateNewInstances(Population *parents, 
			 Population *offspring, 
			 AcyclicOrientedGraph *G, 
			 BoaParams *boaParams)
{
  long  i;
  long  N,M;
  int   k,l;
  int   numberOfNodes;

  int   **group, *groupSize;
  long   *numGroupInstances,*numGroupInstances0,maxNumGroupInstances;
  float **marginalAll, **marginalAllButOne;
  long   *count;  

  char *added;
  int   numAdded;
  char  canAdd;

  char  *x;
  int   where;
  int   position;
  float prob1;

  // set some variables

  numberOfNodes = boaParams->n;
  N             = parents->N;
  M             = offspring->N;

  // allocate the memory for the matrix of groups, their sizes and marginal frequencies
  
  group              = (int**) Calloc(numberOfNodes,sizeof(int*));
  groupSize          = (int*) Calloc(numberOfNodes,sizeof(int));
  numGroupInstances  = (long*) Calloc(numberOfNodes,sizeof(long));
  numGroupInstances0 = (long*) Calloc(numberOfNodes,sizeof(long));
  marginalAll        = (float**) Calloc(numberOfNodes,sizeof(float*));
  marginalAllButOne  = (float**) Calloc(numberOfNodes,sizeof(float*));

  for (k=0; k<numberOfNodes; k++)
    group[k]=(int*) Calloc(numberOfNodes,sizeof(int));

  // allocate the memory for some auxilary array used for ordering vertices topologically

  added = (char*) Malloc(numberOfNodes);

  // create the array of incoming edges numbers for all vartices (with the node
  // on the first position) and the number of them. 
  // + allocate the memory for marginal frequencies

  maxNumGroupInstances = 0;

  for (k=0; k<numberOfNodes; k++)
    {
      // create k-th group (corresponding to the k-th variable)
      
      group[k][0]=k;
      if (G->getNumIn(k)>0)
	memcpy(&(group[k][1]),G->getParentList(k),G->getNumIn(k)*sizeof(int));
      groupSize[k]=G->getNumIn(k)+1;
      
      // compute the number of its instances
      
      numGroupInstances[k] = 1<<groupSize[k];
      numGroupInstances0[k] = numGroupInstances[k]>>1;
      
      // allocate the memory for marginals and counts
      
      marginalAll[k] = (float*) Calloc(numGroupInstances[k],sizeof(float));
      marginalAllButOne[k] = (float*) Calloc(numGroupInstances0[k],sizeof(float));
      
      // update maxNumGroupInstances
      
      if (numGroupInstances[k]>maxNumGroupInstances)
	maxNumGroupInstances=numGroupInstances[k];
    }
  
  count = (long*) Calloc(maxNumGroupInstances,sizeof(long));
  
  // topologically order nodes (and the corresponding groups of the nodes and their parents)
  
  for (k=0; k<numberOfNodes; k++)
    added[k]=0;
  numAdded=0;

  while (numAdded<numberOfNodes)
    {
      for (k=numAdded; k<numberOfNodes; k++)
	{
	  canAdd=1;
	  
	  for (l=1; l<groupSize[k]; l++)
	    if (!added[group[k][l]])
	      canAdd=0;
	  
	  if (canAdd)
	    {
	      added[group[k][0]]=1;
	      if (k!=numAdded)
		{
		  swapPointers((void**) &(group[k]),(void**) &(group[numAdded]));
		  swapInt(&(groupSize[k]), &(groupSize[numAdded]));
		  swapLong(&(numGroupInstances[k]),&(numGroupInstances[numAdded]));
		  swapLong(&(numGroupInstances0[k]),&(numGroupInstances0[numAdded]));
		  swapPointers((void**) &(marginalAll[k]), (void**) &(marginalAll[numAdded]));
		  swapPointers((void**) &(marginalAllButOne[k]), (void**) &(marginalAllButOne[numAdded]));
		}
	      
	      numAdded++;
	    }
	}
    }
  
  // calculate the marginal frequencies for created groups of positions
   
  for (k=0; k<numberOfNodes; k++)
    {
      computeCounts(group[k],groupSize[k],parents,count);
      
      for (l=0; l<numGroupInstances[k]; l++)
	marginalAll[k][l] = (float) count[l]/N;
	     
      for (l=0; l<numGroupInstances0[k]; l++)
	marginalAllButOne[k][l] = marginalAll[k][l]+marginalAll[k][l+numGroupInstances0[k]];
    };
  
  // generate the ith coordinate (chromosome) for all dhildren

  for (i=0; i<M; i++)
    {
      x = offspring->x[i];
      
      for (k=0; k<numberOfNodes; k++)
	{
	  position = group[k][0];
	  
	  if (groupSize[k]==1)
	      prob1 = marginalAll[k][1];
	  else
	    {
	      x[position] = 0;
	      where       = indexedBinaryToInt(x,group[k],groupSize[k]);
	      prob1       = 1-marginalAll[k][where]/marginalAllButOne[k][where];
	    }
	  
	  if (drand()<prob1)
	    x[position]=1;
	  else
	    x[position]=0;
	}
    }

  // free the memory used by marginal frequencies (for groups of bits) and the "count" array

  for (k=0; k<numberOfNodes; k++)
    {
      Free(marginalAll[k]);
      Free(marginalAllButOne[k]);
    };
    
  Free(count);
    
  // free the memory used by the matrix of groups, their sizes and marginal frequencies

  for (k=0; k<numberOfNodes; k++)
    Free(group[k]);

  Free(group);
  Free(groupSize);
  Free(numGroupInstances);
  Free(numGroupInstances0);
  Free(marginalAll);
  Free(marginalAllButOne);

  // get back

  return 0;
};