AdaptiveMonteCarloDecision.C

Go to the documentation of this file.

//* This file is part of the MOOSE framework
//* https://mooseframework.inl.gov
//*
//* All rights reserved, see COPYRIGHT for full restrictions
//* https://github.com/idaholab/moose/blob/master/COPYRIGHT
//*
//* Licensed under LGPL 2.1, please see LICENSE for details
//* https://www.gnu.org/licenses/lgpl-2.1.html
#include "AdaptiveMonteCarloDecision.h"
#include "Sampler.h"
#include "DenseMatrix.h"
#include "AdaptiveMonteCarloUtils.h"
#include "StochasticToolsUtils.h"

#ifdef MOOSE_LIBTORCH_ENABLED
#include "ActiveLearningGPDecision.h"
#endif

registerMooseObject("StochasticToolsApp", AdaptiveMonteCarloDecision);

InputParameters
AdaptiveMonteCarloDecision::validParams()
{
  InputParameters params = GeneralReporter::validParams();
  params.addClassDescription("Generic reporter which decides whether or not to accept a proposed "
                             "sample in Adaptive Monte Carlo type of algorithms.");
  params.addRequiredParam<ReporterName>("output_value",
                                        "Value of the model output from the SubApp.");
  params.addParam<ReporterValueName>(
      "output_required",
      "output_required",
      "Modified value of the model output from this reporter class.");
  params.addParam<ReporterValueName>("inputs", "inputs", "Uncertain inputs to the model.");
  params.addRequiredParam<SamplerName>("sampler", "The sampler object.");
  params.addParam<UserObjectName>("gp_decision", "The Gaussian Process decision reporter.");
  return params;
}

AdaptiveMonteCarloDecision::AdaptiveMonteCarloDecision(const InputParameters & parameters)
  : GeneralReporter(parameters),
    _output_value(isParamValid("gp_decision") ? getReporterValue<std::vector<Real>>("output_value")
                                              : getReporterValue<std::vector<Real>>(
                                                    "output_value", REPORTER_MODE_DISTRIBUTED)),
    _output_required(declareValue<std::vector<Real>>("output_required")),
    _inputs(declareValue<std::vector<std::vector<Real>>>("inputs")),
    _sampler(getSampler("sampler")),
    _ais(dynamic_cast<const AdaptiveImportanceSampler *>(&_sampler)),
    _pss(dynamic_cast<const ParallelSubsetSimulation *>(&_sampler)),
    _check_step(std::numeric_limits<int>::max()),
    _local_comm(_sampler.getLocalComm()),
    _gp_used(isParamValid("gp_decision")),
#ifdef MOOSE_LIBTORCH_ENABLED
    _gp_training_samples(
        _gp_used ? &getUserObject<ActiveLearningGPDecision>("gp_decision").getTrainingSamples()
                 : nullptr)
#else
    _gp_training_samples(nullptr)
#endif
{
#ifndef MOOSE_LIBTORCH_ENABLED
  if (_gp_used)
    paramError("gp_decision", "The 'gp_decision' parameter requires libtorch.");
#endif

  // Check whether the selected sampler is an adaptive sampler or not
  if (!_ais && !_pss)
    paramError("sampler", "The selected sampler is not an adaptive sampler.");

  const auto rows = _sampler.getNumberOfRows();
  const auto cols = _sampler.getNumberOfCols();

  // Initialize the required variables depending upon the type of adaptive Monte Carlo algorithm
  _inputs.resize(cols, std::vector<Real>(rows));
  _prev_val.resize(cols, std::vector<Real>(rows));
  _output_required.resize(rows);
  _prev_val_out.resize(rows);

  if (_ais)
  {
    for (dof_id_type j = 0; j < _sampler.getNumberOfCols(); ++j)
      _prev_val[j][0] = _ais->getInitialValues()[j];
    _output_limit = _ais->getOutputLimit();
  }
  else if (_pss)
  {
    _inputs_sto.resize(cols, std::vector<Real>(_pss->getNumSamplesSub()));
    _inputs_sorted.resize(cols);
    _outputs_sto.resize(_pss->getNumSamplesSub());
    _output_limit = -std::numeric_limits<Real>::max();
  }
}

void
AdaptiveMonteCarloDecision::reinitChain()
{
  const std::vector<Real> & tmp1 = _ais->getInitialValues();
  for (dof_id_type j = 0; j < tmp1.size(); ++j)
    _inputs[j][0] = tmp1[j];
  _prev_val = _inputs;
  _prev_val_out[0] = 1.0;
}

void
AdaptiveMonteCarloDecision::execute()
{
  if (_sampler.getNumberOfLocalRows() == 0 || _check_step == _t_step)
  {
    _check_step = _t_step;
    return;
  }

  /* Decision step to whether or not to accept the proposed sample by the sampler.
     This decision step changes with the type of adaptive Monte Carlo sampling algorithm. */
  if (_ais)
  {
    const Real tmp = _ais->getUseAbsoluteValue() ? std::abs(_output_value[0]) : _output_value[0];

    /* Checking whether a GP surrogate is used. If it is used, importance sampling is not performed
    during the training phase of the GP and all proposed samples are accepted until the training
    phase is completed. Once the training is completed, the importance sampling starts.
    If a GP surrogate is not used, the standard proposal and acceptance/rejection is performed as
    part of the importance sampling. */
    const bool restart_gp = _gp_used && _t_step == *_gp_training_samples;
    const bool output_limit_reached = _gp_used || tmp >= _output_limit;
    if (restart_gp)
      reinitChain();

    _output_required[0] = output_limit_reached ? 1.0 : 0.0;

    if (_t_step <= _ais->getNumSamplesTrain() && !restart_gp)
    {
      /* This is the training phase of the Adaptive Importance Sampling algorithm.
        Here, it is decided whether or not to accept a proposed sample by the
        AdaptiveImportanceSampler.C sampler depending upon the model output_value. */
      _inputs = output_limit_reached
                    ? StochasticTools::reshapeVector(_sampler.getNextLocalRow(), 1, true)
                    : _prev_val;
      if (output_limit_reached)
        _prev_val = _inputs;
      _prev_val_out = _output_required;
    }
    else if (_t_step > _ais->getNumSamplesTrain() && !restart_gp)
    {
      /* This is the sampling phase of the Adaptive Importance Sampling algorithm.
        Here, all proposed samples by the AdaptiveImportanceSampler.C sampler are accepted since
        the importance distribution traning phase is finished. */
      _inputs = StochasticTools::reshapeVector(_sampler.getNextLocalRow(), 1, true);
      _prev_val_out[0] = tmp;
    }
  }
  else if (_pss)
  {
    // Track the current subset
    const unsigned int subset =
        ((_t_step - 1) * _sampler.getNumberOfRows()) / _pss->getNumSamplesSub();
    const unsigned int sub_ind =
        (_t_step - 1) - (_pss->getNumSamplesSub() / _sampler.getNumberOfRows()) * subset;
    const unsigned int offset = sub_ind * _sampler.getNumberOfRows();
    const unsigned int count_max = 1 / _pss->getSubsetProbability();

    DenseMatrix<Real> data_in(_sampler.getNumberOfRows(), _sampler.getNumberOfCols());
    for (dof_id_type ss = _sampler.getLocalRowBegin(); ss < _sampler.getLocalRowEnd(); ++ss)
    {
      const auto data = _sampler.getNextLocalRow();
      for (unsigned int j = 0; j < _sampler.getNumberOfCols(); ++j)
        data_in(ss, j) = data[j];
    }
    _local_comm.sum(data_in.get_values());

    // Get the accepted samples outputs across all the procs from the previous step
    _output_required = (_pss->getUseAbsoluteValue())
                           ? AdaptiveMonteCarloUtils::computeVectorABS(_output_value)
                           : _output_value;
    _local_comm.allgather(_output_required);

    // These are the subsequent subsets which use Markov Chain Monte Carlo sampling scheme
    if (subset > 0)
    {
      if (sub_ind == 0)
      {
        // _output_sorted contains largest po percentile output values
        _output_sorted = AdaptiveMonteCarloUtils::sortOutput(
            _outputs_sto, _pss->getNumSamplesSub(), _pss->getSubsetProbability());
        // _inputs_sorted contains the input values corresponding to the largest po percentile
        // output values
        _inputs_sorted = AdaptiveMonteCarloUtils::sortInput(
            _inputs_sto, _outputs_sto, _pss->getNumSamplesSub(), _pss->getSubsetProbability());
        // Get the subset's intermediate failure threshold values
        _output_limit = AdaptiveMonteCarloUtils::computeMin(_output_sorted);
      }
      // Check whether the number of samples in a Markov chain exceeded the limit
      if (sub_ind % count_max == 0)
      {
        const unsigned int soffset = (sub_ind / count_max) * _sampler.getNumberOfRows();
        // Reinitialize the starting input values for the next set of Markov chains
        for (dof_id_type j = 0; j < _sampler.getNumberOfCols(); ++j)
          _prev_val[j].assign(_inputs_sorted[j].begin() + soffset,
                              _inputs_sorted[j].begin() + soffset + _sampler.getNumberOfRows());
        _prev_val_out.assign(_output_sorted.begin() + soffset,
                             _output_sorted.begin() + soffset + _sampler.getNumberOfRows());
      }
      else
      {
        // Otherwise, use the previously accepted input values to propose the next set of input
        // values
        for (dof_id_type j = 0; j < _sampler.getNumberOfCols(); ++j)
          _prev_val[j].assign(_inputs_sto[j].begin() + offset - _sampler.getNumberOfRows(),
                              _inputs_sto[j].begin() + offset);
        _prev_val_out.assign(_outputs_sto.begin() + offset - _sampler.getNumberOfRows(),
                             _outputs_sto.begin() + offset);
      }
    }

    // Check whether the outputs exceed the subset's intermediate failure threshold value
    for (dof_id_type ss = 0; ss < _sampler.getNumberOfRows(); ++ss)
    {
      // Check whether the outputs exceed the subset's intermediate failure threshold value
      // If so, accept the proposed input values by the Sampler object
      // Otherwise, use the previously accepted input values
      const bool output_limit_reached = _output_required[ss] >= _output_limit;
      for (dof_id_type i = 0; i < _sampler.getNumberOfCols(); ++i)
      {
        _inputs[i][ss] = output_limit_reached ? data_in(ss, i) : _prev_val[i][ss];
        _inputs_sto[i][ss + offset] = _inputs[i][ss];
      }
      if (!output_limit_reached)
        _output_required[ss] = _prev_val_out[ss];
      _outputs_sto[ss + offset] = _output_required[ss];
    }
  }
  // Track the current step
  _check_step = _t_step;
}