DifferentialEvolution::DifferentialEvolution()
: PopulationBasedOptimisationAlgorithm() {
setInitialisingFunctions({{
[this](
const arma::uword numberOfDimensions_,
const arma::mat& initialParameters_) {
population_ = initialParameters_;
return initialParameters_;
},
"Population initialisation"
},
{ [this](
const arma::uword numberOfDimensions_,
const arma::mat& initialParameters_) {
localBestObjectiveValues_.set_size(populationSize_);
localBestObjectiveValues_.fill(std::numeric_limits<double>::infinity());
return initialParameters_;
},
"Local best objective values initialisation"
}
});
setNextParametersFunctions({{
[this](
const arma::uword numberOfDimensions_,
const arma::mat& parameters_,
const arma::rowvec& objectiveValues_,
const arma::rowvec& differences_) {
for (arma::uword n = 0; n < populationSize_; ++n) {
if (objectiveValues_(n) < localBestObjectiveValues_(n)) {
population_.col(n) = parameters_.col(n);
}
}
arma::mat populationCandidates(arma::size(population_));
for (arma::uword n = 0; n < populationSize_; ++n) {
arma::uvec randomIndices = randomPermutationVector(populationSize_, 3);
populationCandidates.col(n) = population_.col(randomIndices(0)) + scalingFactor_ * (population_.col(randomIndices(1)) - population_.col(randomIndices(2)));
}
return populationCandidates;
},
"Differential evolution"
}
});
setScalingFactor(0.5);
}
// -------------------------------------------------------------------------
void MultiMDDAGLearner::parallelRollout(const nor_utils::Args& args, InputData* pData, const string fname, int rsize, GenericClassificationBasedPolicy* policy, PolicyResult* result, const int weakLearnerPostion)
{
vector<AlphaReal> policyError(_shypIter);
vector<InputData*> rollouts(_shypIter,NULL);
// generate rollout
if (_randomNPercent>0)
{
vector<int> randomIndices(_shypIter);
for( int si = 0; si < _shypIter; ++si ) randomIndices[si]=si;
random_shuffle(randomIndices.begin(), randomIndices.end());
int ig = static_cast<int>(static_cast<float>(_shypIter * _randomNPercent) / 100.0);
for( int si = 0; si < ig; ++si )
{
stringstream ss(fname);
// if (si>0)
// {
// ss << fname << "_" << si;
// } else {
// ss << fname;
// }
MDDAGLearner::parallelRollout(args, pData, ss.str(), rsize, policy, result, randomIndices[si]);
InputData* rolloutTrainingData = getRolloutData( args, ss.str() );
if (_verbose)
cout << "---> Rollout size("<< randomIndices[si] << ")" << rolloutTrainingData->getNumExamples() << endl;
rollouts[randomIndices[si]] = rolloutTrainingData;
}
} else {
for( int si = 0; si < _shypIter; ++si )
{
stringstream ss(fname);
// if (si>0)
// {
// ss << fname << "_" << si;
// } else {
// ss << fname;
// }
MDDAGLearner::parallelRollout(args, pData, ss.str(), rsize, policy, result, si);
InputData* rolloutTrainingData = getRolloutData( args, ss.str() );
if (_verbose)
cout << "---> Rollout size("<< si << ")" << rolloutTrainingData->getNumExamples() << endl;
rollouts[si] = rolloutTrainingData;
}
}
// update policy
int numOfUpdatedPolicy = 0;
for( int si = 0; si < _shypIter; ++si )
{
if ((rollouts[si]==NULL) || (rollouts[si]->getNumExamples()<=2)) continue;
policyError[si] = _policy->trainpolicy( rollouts[si], _baseLearnerName, _trainingIter, si );
if (_verbose)
cout << "--> Policy error: pos: " << si << "\t error:\t" << setprecision (4) << policyError[si] << endl;
numOfUpdatedPolicy++;
}
if (_verbose)
cout << "--> Number of updated policy" << numOfUpdatedPolicy << endl << flush;
//release rolouts
for( int si = 0; si < _shypIter; ++si )
{
if (rollouts[si]) delete rollouts[si];
}
}