public override EvaluationResults EvaluateModelOnData(Converter <Leaf, SufficientStatistics> predictorMap, Converter <Leaf, SufficientStatistics> targetMap)
{
int[] realFisherCounts = ModelScorer.PhyloTree.FisherCounts(predictorMap, targetMap);
int tt = realFisherCounts[(int)TwoByTwo.ParameterIndex.TT];
int tf = realFisherCounts[(int)TwoByTwo.ParameterIndex.TF];
int ft = realFisherCounts[(int)TwoByTwo.ParameterIndex.FT];
int ff = realFisherCounts[(int)TwoByTwo.ParameterIndex.FF];
int[] fisherCountsPred = new int[] { tt, ft, tf, ff }; //ModelScorer.PhyloTree.FisherCounts(targetMap, predictorMap);
int[] fisherCountsTarg = realFisherCounts;
#if NAIVE_EQUILIBRIUM
//USE THIS FOR BACKWARDS COMPATABILITY
int[] tempCountsPred = ModelScorer.PhyloTree.CountsOfLeaves(predictorMap);
int[] tempCountsTarg = ModelScorer.PhyloTree.CountsOfLeaves(targetMap);
fisherCountsPred = tempCountsPred;
fisherCountsTarg = tempCountsTarg;
#endif
bool predIsInvariant, targIsInvariant;
Score nullScorePred = ComputeSingleVariableScore(targetMap, predictorMap, (DistributionDiscreteSingleVariable)NullDistns[0], fisherCountsPred, out predIsInvariant);
Score nullScoreTarg = ComputeSingleVariableScore(predictorMap, targetMap, (DistributionDiscreteSingleVariable)NullDistns[1], fisherCountsTarg, out targIsInvariant);
List <Score> nullScores = new List <Score>(new Score[] { nullScorePred, nullScoreTarg });
OptimizationParameterList initParams = ((DistributionDiscreteJoint)AltDistn).GenerateInitialParams(nullScorePred.OptimizationParameters, nullScoreTarg.OptimizationParameters);
Score jointScore;
if (predIsInvariant || targIsInvariant) // cannot compute parameters in this case. They come directly from the single variable params
{
double jointLL = nullScorePred.Loglikelihood + nullScoreTarg.Loglikelihood;
jointScore = Score.GetInstance(jointLL, initParams, AltDistn);
}
else
{
MessageInitializerDiscrete altMessageInitializer = MessageInitializerDiscrete.GetInstance(CreateJointMap(predictorMap, targetMap), (DistributionDiscreteJoint)AltDistn, initParams, ModelScorer.PhyloTree.LeafCollection);
jointScore = ModelScorer.MaximizeLikelihood(altMessageInitializer);
}
EvaluationResults evalResults = EvaluationResultsDiscrete.GetInstance(this, nullScores, jointScore, realFisherCounts, ChiSquareDegreesOfFreedom);
return(evalResults);
}
private Score ComputeConditionalVariableScore(
Converter <Leaf, SufficientStatistics> predictorMap,
Converter <Leaf, SufficientStatistics> targetMap,
Score nullScore,
int[] fisherCounts)
{
int tt = fisherCounts[(int)TwoByTwo.ParameterIndex.TT];
int tf = fisherCounts[(int)TwoByTwo.ParameterIndex.TF];
int ft = fisherCounts[(int)TwoByTwo.ParameterIndex.FT];
int sum = SpecialFunctions.Sum(fisherCounts);
Score altScore;
if (tt + ft == sum || tt + ft == 0) // target is always true or false
{
bool isNaN = sum == 0;
OptimizationParameterList altParamList = AltDistn.GetParameters();
altParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Predictor1].Value = 0;
altParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Predictor2].Value = 0;
altParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Lambda].Value = isNaN ? double.NaN : 0;
altParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Equilibrium].Value = (double)(tt + ft) / sum;
altScore = Score.GetInstance(isNaN ? double.NaN : 0, altParamList, AltDistn);
}
else if (tt + tf == 0 || tt + tf == sum) // predictor is always true or false
{
OptimizationParameterList nullParamList = nullScore.OptimizationParameters;
OptimizationParameterList altParamList = AltDistn.GetParameters();
altParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Predictor1].Value = nullParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Predictor1].Value;
altParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Predictor2].Value = nullParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Predictor2].Value;
altParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Lambda].Value = nullParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Lambda].Value;
altParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Equilibrium].Value = nullParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Equilibrium].Value;
altScore = Score.GetInstance(nullScore.Loglikelihood, altParamList, AltDistn);
}
else // compute ML using ModelScorer
{
MessageInitializerDiscrete altMessageInitializer = MessageInitializerDiscrete.GetInstance(predictorMap, targetMap, (DistributionDiscreteConditional)AltDistn, nullScore.OptimizationParameters, ModelScorer.PhyloTree.LeafCollection);
altScore = ModelScorer.MaximizeLikelihood(altMessageInitializer);
}
return(altScore);
}
//public static Converter<Leaf, SufficientStatistics> TESTPRED, TESTTARG;
//public static EvaluationResults TESTEVALRESULTS;
public override EvaluationResults EvaluateModelOnDataGivenParams(
Converter <Leaf, SufficientStatistics> predictorMap, Converter <Leaf, SufficientStatistics> targetMap, EvaluationResults previousResults)
{
int[] fisherCounts = ModelScorer.PhyloTree.FisherCounts(predictorMap, targetMap);
int targNullIdx = _includePredictorInScore ? 1 : 0;
OptimizationParameterList nullParamsTarg = previousResults.NullScores[targNullIdx].OptimizationParameters;
MessageInitializerDiscrete nullMessageInitializerTarg = MessageInitializerDiscrete.GetInstance(predictorMap, targetMap,
NullDistn, new int[0], ModelScorer.PhyloTree.LeafCollection);
double nullLLTarg = ModelScorer.ComputeLogLikelihoodModelGivenData(nullMessageInitializerTarg, nullParamsTarg);
Score nullScoreTarg = Score.GetInstance(nullLLTarg, nullParamsTarg, NullDistn);
OptimizationParameterList altParams = previousResults.AltScore.OptimizationParameters;
MessageInitializerDiscrete altMessageInitializer = MessageInitializerDiscrete.GetInstance(predictorMap, targetMap,
(DistributionDiscreteConditional)AltDistn, new int[0], ModelScorer.PhyloTree.LeafCollection);
double condLL = ModelScorer.ComputeLogLikelihoodModelGivenData(altMessageInitializer, altParams);
Score altScore = Score.GetInstance(condLL, altParams, AltDistn);
List <Score> nullScores = new List <Score>();
if (_includePredictorInScore)
{
OptimizationParameterList nullParamsPred = previousResults.NullScores[0].OptimizationParameters;
MessageInitializerDiscrete nullMessageInitializerPred = MessageInitializerDiscrete.GetInstance(targetMap, predictorMap,
NullDistn, new int[0], ModelScorer.PhyloTree.LeafCollection);
double nullLLPred = ModelScorer.ComputeLogLikelihoodModelGivenData(nullMessageInitializerPred, nullParamsPred);
Score nullScorePred = Score.GetInstance(nullLLPred, nullParamsPred, NullDistn);
nullScores.Add(nullScorePred);
// conditional model altScore doesn't include predLL. If we're here, we want to add it to make it comparable to joint or reverseConditional
altScore = Score.GetInstance(altScore.Loglikelihood + nullScorePred.Loglikelihood, altScore.OptimizationParameters, altScore.Distribution);
}
nullScores.Add(nullScoreTarg);
EvaluationResults evalResults = EvaluationResultsDiscrete.GetInstance(this, nullScores, altScore, fisherCounts, ChiSquareDegreesOfFreedom);
return(evalResults);
}
public override EvaluationResults EvaluateModelOnDataGivenParams(Converter <Leaf, SufficientStatistics> predictorMap, Converter <Leaf, SufficientStatistics> targetMap, EvaluationResults previousResults)
{
int[] fisherCounts = ModelScorer.PhyloTree.FisherCounts(predictorMap, targetMap);
OptimizationParameterList nullParamsTarg = previousResults.NullScores[1].OptimizationParameters;
MessageInitializerDiscrete nullMessageInitializerTarg = MessageInitializerDiscrete.GetInstance(predictorMap, targetMap, (DistributionDiscreteSingleVariable)NullDistns[0], fisherCounts, ModelScorer.PhyloTree.LeafCollection);
double nullLLTarg = ModelScorer.ComputeLogLikelihoodModelGivenData(nullMessageInitializerTarg, nullParamsTarg);
Score nullScoreTarg = Score.GetInstance(nullLLTarg, nullParamsTarg, previousResults.NullScores[1].Distribution);
OptimizationParameterList nullParamsPred = previousResults.NullScores[0].OptimizationParameters;
MessageInitializerDiscrete nullMessageInitializerPred = MessageInitializerDiscrete.GetInstance(targetMap, predictorMap, (DistributionDiscreteSingleVariable)NullDistns[1], fisherCounts, ModelScorer.PhyloTree.LeafCollection);
double nullLLPred = ModelScorer.ComputeLogLikelihoodModelGivenData(nullMessageInitializerPred, nullParamsPred);
Score nullScorePred = Score.GetInstance(nullLLPred, nullParamsPred, previousResults.NullScores[0].Distribution);
List <Score> nullScores = new List <Score>(new Score[] { nullScorePred, nullScoreTarg });
OptimizationParameterList altParams = previousResults.AltScore.OptimizationParameters;
double altLL;
if (((DistributionDiscreteJoint)AltDistn).ParametersCannotBeEvaluated(altParams))
{
// we'll get here only if one of the variables is always (or never) true. In this case, the variables must be independent.
altLL = nullLLTarg + nullLLPred;
}
else
{
MessageInitializerDiscrete altMessageInitializer = MessageInitializerDiscrete.GetInstance(CreateJointMap(predictorMap, targetMap), (DistributionDiscreteJoint)AltDistn, fisherCounts, ModelScorer.PhyloTree.LeafCollection);
altLL = ModelScorer.ComputeLogLikelihoodModelGivenData(altMessageInitializer, altParams);
}
Score altScore = Score.GetInstance(altLL, altParams, previousResults.AltScore.Distribution);
EvaluationResults evalResults = EvaluationResultsDiscrete.GetInstance(this, nullScores, altScore, fisherCounts, ChiSquareDegreesOfFreedom);
return(evalResults);
}
public override Dictionary <string, SufficientStatistics> GenerateRandomMapping(Dictionary <string, SufficientStatistics> realCaseIdToNonMissingValue, ref Random random)
{
//!!!!put check in to make sure ISufficientSTatistics is reall BooleanStatistics
Converter <Leaf, SufficientStatistics> leafToDistnClassFunction = PhyloDDriver.CreateSufficientStatisticsMap(realCaseIdToNonMissingValue);
PhyloTree tree = _modelScorer.PhyloTree;
MessageInitializer messageInitializer = MessageInitializerDiscrete.GetInstance(leafToDistnClassFunction, _discreteDistribution, new int[] { 1, 1 }, tree.LeafCollection);
Score score = _modelScorer.MaximizeLikelihood(messageInitializer);
double percentNonMissing = (double)tree.CountOfNonMissingLeaves(realCaseIdToNonMissingValue) /
(double)SpecialFunctions.Count(tree.LeafCollection);
double equilibrium = score.OptimizationParameters[(int)DistributionDiscreteConditional.ParameterIndex.Equilibrium].Value;
double lambda = score.OptimizationParameters[(int)DistributionDiscreteConditional.ParameterIndex.Lambda].Value;
Dictionary <string, BooleanStatistics> randomCaseIdToNonMissingValue = tree.EvolveBinaryTree(equilibrium, lambda, 1 - percentNonMissing, ref random);
Dictionary <string, SufficientStatistics> converted;
SpecialFunctions.ConvertDictionaryToBaseClasses(randomCaseIdToNonMissingValue, out converted);
return(converted);
}
protected Score ComputeSingleVariableScore(
Converter <Leaf, SufficientStatistics> predictorMap,
Converter <Leaf, SufficientStatistics> targetMap,
DistributionDiscreteSingleVariable nullDistn,
int[] fisherCounts,
out bool variableIsInvariant)
{
MessageInitializerDiscrete nullMessageInitializer =
MessageInitializerDiscrete.GetInstance(predictorMap, targetMap, nullDistn, fisherCounts, ModelScorer.PhyloTree.LeafCollection);
double p = (double)TwoByTwo.GetRightSum(fisherCounts) / SpecialFunctions.Sum(fisherCounts);
Score nullScore;
if (TryGetSingleVariableScoreFromCounts(nullMessageInitializer, p, out nullScore))
{
variableIsInvariant = true;
}
else
{
variableIsInvariant = false;
nullScore = ModelScorer.MaximizeLikelihood(nullMessageInitializer);
}
return(nullScore);
}
private bool TryGetSingleVariableScoreFromCounts(MessageInitializerDiscrete singleVariableMessageInitializer, double pVar, out Score score)
{
if (pVar == 1 || pVar == 0)
{
OptimizationParameterList nullParamList = singleVariableMessageInitializer.DiscreteDistribution.GetParameters();
nullParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Lambda].Value = 0;
nullParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Equilibrium].Value = pVar;
score = Score.GetInstance(0, nullParamList, singleVariableMessageInitializer.DiscreteDistribution);
return(true);
}
else if (double.IsNaN(pVar))
{
OptimizationParameterList nullParamList = singleVariableMessageInitializer.DiscreteDistribution.GetParameters();
nullParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Lambda].Value = double.NaN;
nullParamList[(int)DistributionDiscreteSingleVariable.ParameterIndex.Equilibrium].Value = double.NaN;
score = Score.GetInstance(double.NaN, nullParamList, singleVariableMessageInitializer.DiscreteDistribution);
return(true);
}
else
{
score = null;
return(false);
}
}
public override EvaluationResults EvaluateModelOnData(Converter <Leaf, SufficientStatistics> predictorMap, Converter <Leaf, SufficientStatistics> targetMap)
{
EvaluationResults evalResults;
int[] fisherCounts = ModelScorer.PhyloTree.FisherCounts(predictorMap, targetMap);
int[] realFisherCounts = fisherCounts; // for compatability when NAIVE_EQUILIBRIUM is set
#if NAIVE_EQUILIBRIUM
//USE THIS FOR BACKWARDS COMPATABILITY
int[] tempCounts = ModelScorer.PhyloTree.CountsOfLeaves(targetMap);
fisherCounts = tempCounts;
#endif
//MessageInitializerDiscrete nullMessageInitializer = MessageInitializerDiscrete.GetInstance(predictorMap, targetMap, NullDistn, fisherCounts, ModelScorer.PhyloTree.LeafCollection);
//if (TryShortCutFromCounts(realFisherCounts, nullMessageInitializer, out evalResults))
//{
// return evalResults;
//}
//Score nullScore = ModelScorer.MaximizeLikelihood(nullMessageInitializer);
bool isInvariant;
Score nullScoreTarg = ComputeSingleVariableScore(predictorMap, targetMap, NullDistn, fisherCounts, out isInvariant);
Score altScore = ComputeConditionalVariableScore(predictorMap, targetMap, nullScoreTarg, fisherCounts);
//(realFisherCounts, nullScoreTarg, out evalResults))
//{
// return evalResults;
//}
//MessageInitializerDiscrete altMessageInitializer = MessageInitializerDiscrete.GetInstance(predictorMap, targetMap, (DistributionDiscreteConditional)AltDistn, nullScore.OptimizationParameters, ModelScorer.PhyloTree.LeafCollection);
//Score condScore = ModelScorer.MaximizeLikelihood(altMessageInitializer);
List <Score> nullScores = new List <Score>();
if (_includePredictorInScore)
{
int[] predFisherCounts = new int[] { realFisherCounts[0], realFisherCounts[2], realFisherCounts[1], realFisherCounts[3] };
Score predNullScore = ComputeSingleVariableScore(targetMap, predictorMap, NullDistn, predFisherCounts, out isInvariant);
nullScores.Add(predNullScore);
// conditional model altScore doesn't include predLL. If we're here, we want to add it to make it comparable to joint or reverseConditional
altScore = Score.GetInstance(altScore.Loglikelihood + predNullScore.Loglikelihood, altScore.OptimizationParameters, altScore.Distribution);
}
nullScores.Add(nullScoreTarg);
evalResults = EvaluationResultsDiscrete.GetInstance(this, nullScores, altScore, realFisherCounts, ChiSquareDegreesOfFreedom);
#if DEBUG
MessageInitializerDiscrete nullMessageInitializer = MessageInitializerDiscrete.GetInstance(predictorMap, targetMap, NullDistn, fisherCounts, ModelScorer.PhyloTree.LeafCollection);
MessageInitializerDiscrete altMessageInitializer = MessageInitializerDiscrete.GetInstance(predictorMap, targetMap, (DistributionDiscreteConditional)AltDistn, nullScoreTarg.OptimizationParameters, ModelScorer.PhyloTree.LeafCollection);
double nullLL = ModelScorer.ComputeLogLikelihoodModelGivenData(nullMessageInitializer, nullScoreTarg.OptimizationParameters);
double altLL = ModelScorer.ComputeLogLikelihoodModelGivenData(altMessageInitializer, altScore.OptimizationParameters);
if (_includePredictorInScore)
{
int[] predFisherCounts = new int[] { realFisherCounts[0], realFisherCounts[2], realFisherCounts[1], realFisherCounts[3] };
MessageInitializerDiscrete nullMessageInitializerPred = MessageInitializerDiscrete.GetInstance(targetMap, predictorMap, NullDistn, predFisherCounts, ModelScorer.PhyloTree.LeafCollection);
double nullLLPred = ModelScorer.ComputeLogLikelihoodModelGivenData(nullMessageInitializerPred, nullScores[0].OptimizationParameters);
altLL += nullLLPred;
}
EvaluationResults evalResults2 = EvaluateModelOnDataGivenParams(predictorMap, targetMap, evalResults);
double eps = 1E-10;
Debug.Assert(ComplexNumber.ApproxEqual(nullLL, nullScoreTarg.Loglikelihood, eps));
Debug.Assert(ComplexNumber.ApproxEqual(altLL, altScore.Loglikelihood, eps));
Debug.Assert(ComplexNumber.ApproxEqual(evalResults.NullLL, evalResults2.NullLL, eps) && ComplexNumber.ApproxEqual(evalResults.AltLL, evalResults2.AltLL, eps), "In ModelEvaluatorCond, results of maximizing LL and computing LL from same params are not the same.");
#endif
return(evalResults);
}