potential.java

是实现关系型贝叶斯网络一中机器学习算法

package rmn;

import java.util.*;

public class Potential {

  // potential factory, used for getting potential weights
  PotentialFactory m_pf;
  
  // attached variables
  Variable[] m_vars;

  // array of messages, one message for each variable
  double[][] m_msgPotToVar;
  double[][] m_oldPotToVar;

  public Potential(PotentialFactory pf)
  {
    m_pf = pf;
  }

  public void allocateMessages()
  {
    m_msgPotToVar = new double[m_vars.length][];
    m_oldPotToVar = new double[m_vars.length][];
    for (int i = 0; i < m_vars.length; i++) {
      m_msgPotToVar[i] = new double[m_vars[i].m_nCard];
      m_oldPotToVar[i] = new double[m_vars[i].m_nCard];
    }
  }

  public void deallocateMessages()
  {
    m_msgPotToVar = null;
    m_oldPotToVar = null;
  }


  public void attachVariables(Variable[] vars)
  {
    //    assert vars.length == m_pf.size() : vars.length;

    //    m_vars = new Variable[m_pf.size()];
    m_vars = new Variable[vars.length];
    for (int i = 0; i < m_vars.length; i++)
      m_vars[i] = vars[i];
  }
  

  public double[] getMessage(Variable var)
  {
    for (int i = 0; i < m_vars.length; i++) 
      if (m_vars[i].compareTo(var) == 0)
        return m_msgPotToVar[i];

    return null;
  }

  public double[] getOldMessage(Variable var)
  {
    for (int i = 0; i < m_vars.length; i++) 
      if (m_vars[i].compareTo(var) == 0)
        return m_oldPotToVar[i];

    return null;
  }

  public double[] getMessage(int var)
  {
    return m_msgPotToVar[var];
  }

  public double[] getOldMessage(int var)
  {
    return m_oldPotToVar[var];
  }

  public void setMessage(int var, double[] msg)
  {
    assert msg.length == m_vars[var].m_nCard: msg.length;

    System.arraycopy(msg, 0, m_msgPotToVar[var], 0, msg.length);
  }

  public void initMessages()
  {
    for (int i = 0; i < m_vars.length; i++)
      Arrays.fill(m_msgPotToVar[i], 1);
  }

  public Matrix newWeightMatrix()
  {
    Matrix m = m_pf.newWeightMatrix();
    m.fill(1);

    return m;
  }

  public Matrix getWeightMatrix()
  {
    return m_pf.getWeightMatrix();
  }

  public void backupMessages()
  {
    for (int i = 0; i < m_vars.length; i++)
      for (int j = 0; j < m_vars[i].m_nCard; j++)
        m_oldPotToVar[i][j] = m_msgPotToVar[i][j];
  }


  public boolean sendMsgToVars(boolean bMaximize)
  {
    boolean bConverged = true;

    for (int v = 0; v < m_vars.length; v++) {
      Variable var = m_vars[v];
      // absorb variable messages
      Matrix temp1 = newWeightMatrix();
      for (int vv = 0; vv < m_vars.length; vv++) {
	if (vv != v) {
	  Variable varvar = m_vars[vv];
	  temp1.dotProduct(varvar.getOldMessage(this), vv);
	}
      }
      temp1.dotProduct(getWeightMatrix());
      double[] temp2 = temp1.marginalize(v, bMaximize);
      MathUtils.normalize(temp2);
      setMessage(v, temp2);
      // need more time?
      if (!MathUtils.approxeq(getMessage(v), getOldMessage(v),
			      FactorGraph.TOLERANCE))
	bConverged = false;
    }
    
    return bConverged;
  }

}