ROL_PH_RiskObjective.hpp

Go to the documentation of this file.

// @HEADER
// *****************************************************************************
//               Rapid Optimization Library (ROL) Package
//
// Copyright 2014 NTESS and the ROL contributors.
// SPDX-License-Identifier: BSD-3-Clause
// *****************************************************************************
// @HEADER
 
#ifndef PH_RISKOBJECTIVE_H
#define PH_RISKOBJECTIVE_H
 
#include "ROL_Objective.hpp"
#include "ROL_RiskMeasureFactory.hpp"
 
namespace ROL {
 
template <class Real>
class PH_RiskObjective : public Objective<Real> {
private:
  const Ptr<Objective<Real>> obj_;
  Ptr<ExpectationQuad<Real>> quad_;
 
  bool isValueComputed_;
  Real val_;
 
  bool isGradientInitialized_;
  bool isGradientComputed_;
  Ptr<Vector<Real>> g_;
 
  void getValue(const Vector<Real> &x, Real &tol) {
    if (!isValueComputed_) {
      val_ = obj_->value(x,tol);
      isValueComputed_ = true;
    }
  }
 
  void getGradient(const Vector<Real> &x, Real &tol) {
    if (!isGradientInitialized_) {
      g_ = x.dual().clone();
      isGradientInitialized_ = true;
    }
    if (!isGradientComputed_) {
      obj_->gradient(*g_,x,tol);
      isGradientComputed_ = true;
    }
  }
 
  Ptr<const Vector<Real>> getConstVector(const Vector<Real> &x) const {
    const RiskVector<Real> &xrv = dynamic_cast<const RiskVector<Real>&>(x);
    return xrv.getVector();
  }
 
  Ptr<Vector<Real>> getVector(Vector<Real> &x) const {
    RiskVector<Real> &xrv = dynamic_cast<RiskVector<Real>&>(x);
    return xrv.getVector();
  }
 
  Ptr<const std::vector<Real>> getConstStat(const Vector<Real> &x) const {
    const RiskVector<Real> &xrv = dynamic_cast<const RiskVector<Real>&>(x);
    Ptr<const std::vector<Real>> xstat = xrv.getStatistic();
    if (xstat == nullPtr) {
      xstat = makePtr<const std::vector<Real>>(0);
    }
    return xstat;
  }
 
  Ptr<std::vector<Real>> getStat(Vector<Real> &x) const {
    RiskVector<Real> &xrv = dynamic_cast<RiskVector<Real>&>(x);
    Ptr<std::vector<Real>> xstat = xrv.getStatistic();
    if (xstat == nullPtr) {
      xstat = makePtr<std::vector<Real>>(0);
    }
    return xstat;
  }
 
public:
 
  PH_RiskObjective(const Ptr<Objective<Real>> &obj,
                        ParameterList              &parlist)
    : obj_(obj),
      isValueComputed_(false),
      isGradientInitialized_(false),
      isGradientComputed_(false) {
    std::string risk = parlist.sublist("SOL").sublist("Risk Measure").get("Name","CVaR");
    ERiskMeasure ed = StringToERiskMeasure(risk);
    switch(ed) {
      case RISKMEASURE_CVAR:
             quad_ = makePtr<QuantileQuadrangle<Real>>(parlist);          break;
      case RISKMEASURE_MOREAUYOSIDACVAR:
             quad_ = makePtr<MoreauYosidaCVaR<Real>>(parlist);            break;
      case RISKMEASURE_GENMOREAUYOSIDACVAR:
             quad_ = makePtr<GenMoreauYosidaCVaR<Real>>(parlist);         break;
      case RISKMEASURE_LOGEXPONENTIAL:
             quad_ = makePtr<LogExponentialQuadrangle<Real>>(parlist);    break;
      case RISKMEASURE_SAFETYMARGIN:
             quad_ = makePtr<MeanVarianceQuadrangle<Real>>(parlist);      break;
      case RISKMEASURE_TRUNCATEDMEAN:
             quad_ = makePtr<TruncatedMeanQuadrangle<Real>>(parlist);     break;
      case RISKMEASURE_LOGQUANTILE:
             quad_ = makePtr<LogQuantileQuadrangle<Real>>(parlist);       break;
      case RISKMEASURE_SMOOTHEDWORSTCASE:
             quad_ = makePtr<SmoothedWorstCaseQuadrangle<Real>>(parlist); break;
//      case RISKMEASURE_CHI2DIVERGENCE:
//             return makePtr<Chi2Divergence<Real>>(parlist);
//      case RISKMEASURE_KLDIVERGENCE:
//             return makePtr<KLDivergence<Real>>(parlist);
      default:
        ROL_TEST_FOR_EXCEPTION(true,std::invalid_argument,
                                   "Invalid risk measure type " << risk << "!");
    }
  }
 
  void update( const Vector<Real> &x, bool flag = true, int iter = -1 ) {
    Ptr<const Vector<Real>> xvec = getConstVector(x);
    obj_->update(*xvec,flag,iter);
    isValueComputed_    = false;
    isGradientComputed_ = false;
  }
 
  Real value( const Vector<Real> &x, Real &tol ) {
    Ptr<const Vector<Real>> xvec = getConstVector(x);
    Ptr<const std::vector<Real>> xstat = getConstStat(x);
    getValue(*xvec,tol);
    Real reg = quad_->regret(val_-(*xstat)[0],0);
    return (*xstat)[0] + reg;
  }
 
  void gradient( Vector<Real> &g, const Vector<Real> &x, Real &tol ) {
    Ptr<Vector<Real>>            gvec  = getVector(g);
    Ptr<std::vector<Real>>       gstat = getStat(g);
    Ptr<const Vector<Real>>      xvec  = getConstVector(x);
    Ptr<const std::vector<Real>> xstat = getConstStat(x);
    getValue(*xvec,tol);
    Real reg = quad_->regret(val_-(*xstat)[0],1);
    getGradient(*xvec,tol);
    gvec->set(*g_); gvec->scale(reg);
    (*gstat)[0] = static_cast<Real>(1)-reg;
  }
 
  void hessVec( Vector<Real> &hv, const Vector<Real> &v, const Vector<Real> &x, Real &tol ) {
    Ptr<Vector<Real>>            hvec  = getVector(hv);
    Ptr<std::vector<Real>>       hstat = getStat(hv);
    Ptr<const Vector<Real>>      vvec  = getConstVector(v);
    Ptr<const std::vector<Real>> vstat = getConstStat(v);
    Ptr<const Vector<Real>>      xvec  = getConstVector(x);
    Ptr<const std::vector<Real>> xstat = getConstStat(x);
    getValue(*xvec,tol);
    Real reg1 = quad_->regret(val_-(*xstat)[0],1);
    Real reg2 = quad_->regret(val_-(*xstat)[0],2);
    getGradient(*xvec,tol);
    //Real gv   = vvec->dot(g_->dual());
    Real gv   = vvec->apply(*g_);
    obj_->hessVec(*hvec,*vvec,*xvec,tol);
    hvec->scale(reg1); hvec->axpy(reg2*(gv-(*vstat)[0]),*g_);
    (*hstat)[0] = reg2*((*vstat)[0]-gv);
  }
 
  void setParameter(const std::vector<Real> &param) {
    obj_->setParameter(param);
    Objective<Real>::setParameter(param);
  }
 
};
 
}
#endif