| Line |
Branch |
Exec |
Source |
| 1 |
|
|
/////////////////////////////////////////////////////////////////////////////// |
| 2 |
|
|
// BSD 3-Clause License |
| 3 |
|
|
// |
| 4 |
|
|
// Copyright (C) 2019-2022, LAAS-CNRS, University of Edinburgh |
| 5 |
|
|
// Heriot-Watt University |
| 6 |
|
|
// Copyright note valid unless otherwise stated in individual files. |
| 7 |
|
|
// All rights reserved. |
| 8 |
|
|
/////////////////////////////////////////////////////////////////////////////// |
| 9 |
|
|
|
| 10 |
|
|
#ifdef CROCODDYL_WITH_MULTITHREADING |
| 11 |
|
|
#include <omp.h> |
| 12 |
|
|
#endif // CROCODDYL_WITH_MULTITHREADING |
| 13 |
|
|
|
| 14 |
|
|
#include "crocoddyl/core/solvers/fddp.hpp" |
| 15 |
|
|
|
| 16 |
|
|
namespace crocoddyl { |
| 17 |
|
|
|
| 18 |
|
✗ |
SolverFDDP::SolverFDDP(std::shared_ptr<ShootingProblem> problem) |
| 19 |
|
✗ |
: SolverDDP(problem), dg_(0), dq_(0), dv_(0), th_acceptnegstep_(2) {} |
| 20 |
|
|
|
| 21 |
|
✗ |
SolverFDDP::~SolverFDDP() {} |
| 22 |
|
|
|
| 23 |
|
✗ |
bool SolverFDDP::solve(const std::vector<Eigen::VectorXd>& init_xs, |
| 24 |
|
|
const std::vector<Eigen::VectorXd>& init_us, |
| 25 |
|
|
const std::size_t maxiter, const bool is_feasible, |
| 26 |
|
|
const double init_reg) { |
| 27 |
|
✗ |
START_PROFILER("SolverFDDP::solve"); |
| 28 |
|
✗ |
if (problem_->is_updated()) { |
| 29 |
|
✗ |
resizeData(); |
| 30 |
|
|
} |
| 31 |
|
✗ |
xs_try_[0] = |
| 32 |
|
✗ |
problem_->get_x0(); // it is needed in case that init_xs[0] is infeasible |
| 33 |
|
✗ |
setCandidate(init_xs, init_us, is_feasible); |
| 34 |
|
|
|
| 35 |
|
✗ |
if (std::isnan(init_reg)) { |
| 36 |
|
✗ |
preg_ = reg_min_; |
| 37 |
|
✗ |
dreg_ = reg_min_; |
| 38 |
|
|
} else { |
| 39 |
|
✗ |
preg_ = init_reg; |
| 40 |
|
✗ |
dreg_ = init_reg; |
| 41 |
|
|
} |
| 42 |
|
✗ |
was_feasible_ = false; |
| 43 |
|
|
|
| 44 |
|
✗ |
bool recalcDiff = true; |
| 45 |
|
✗ |
for (iter_ = 0; iter_ < maxiter; ++iter_) { |
| 46 |
|
|
while (true) { |
| 47 |
|
|
try { |
| 48 |
|
✗ |
computeDirection(recalcDiff); |
| 49 |
|
✗ |
} catch (std::exception& e) { |
| 50 |
|
✗ |
recalcDiff = false; |
| 51 |
|
✗ |
increaseRegularization(); |
| 52 |
|
✗ |
if (preg_ == reg_max_) { |
| 53 |
|
✗ |
return false; |
| 54 |
|
|
} else { |
| 55 |
|
✗ |
continue; |
| 56 |
|
|
} |
| 57 |
|
|
} |
| 58 |
|
✗ |
break; |
| 59 |
|
|
} |
| 60 |
|
✗ |
updateExpectedImprovement(); |
| 61 |
|
|
|
| 62 |
|
|
// We need to recalculate the derivatives when the step length passes |
| 63 |
|
✗ |
recalcDiff = false; |
| 64 |
|
✗ |
for (std::vector<double>::const_iterator it = alphas_.begin(); |
| 65 |
|
✗ |
it != alphas_.end(); ++it) { |
| 66 |
|
✗ |
steplength_ = *it; |
| 67 |
|
|
|
| 68 |
|
|
try { |
| 69 |
|
✗ |
dV_ = tryStep(steplength_); |
| 70 |
|
✗ |
} catch (std::exception& e) { |
| 71 |
|
✗ |
continue; |
| 72 |
|
|
} |
| 73 |
|
✗ |
expectedImprovement(); |
| 74 |
|
✗ |
dVexp_ = steplength_ * (d_[0] + 0.5 * steplength_ * d_[1]); |
| 75 |
|
|
|
| 76 |
|
✗ |
if (dVexp_ >= 0) { // descend direction |
| 77 |
|
✗ |
if (std::abs(d_[0]) < th_grad_ || dV_ > th_acceptstep_ * dVexp_) { |
| 78 |
|
✗ |
was_feasible_ = is_feasible_; |
| 79 |
|
✗ |
setCandidate(xs_try_, us_try_, (was_feasible_) || (steplength_ == 1)); |
| 80 |
|
✗ |
cost_ = cost_try_; |
| 81 |
|
✗ |
recalcDiff = true; |
| 82 |
|
✗ |
break; |
| 83 |
|
|
} |
| 84 |
|
|
} else { // reducing the gaps by allowing a small increment in the cost |
| 85 |
|
|
// value |
| 86 |
|
✗ |
if (!is_feasible_ && dV_ > th_acceptnegstep_ * dVexp_) { |
| 87 |
|
✗ |
was_feasible_ = is_feasible_; |
| 88 |
|
✗ |
setCandidate(xs_try_, us_try_, (was_feasible_) || (steplength_ == 1)); |
| 89 |
|
✗ |
cost_ = cost_try_; |
| 90 |
|
✗ |
recalcDiff = true; |
| 91 |
|
✗ |
break; |
| 92 |
|
|
} |
| 93 |
|
|
} |
| 94 |
|
|
} |
| 95 |
|
|
|
| 96 |
|
✗ |
if (steplength_ > th_stepdec_) { |
| 97 |
|
✗ |
decreaseRegularization(); |
| 98 |
|
|
} |
| 99 |
|
✗ |
if (steplength_ <= th_stepinc_) { |
| 100 |
|
✗ |
increaseRegularization(); |
| 101 |
|
✗ |
if (preg_ == reg_max_) { |
| 102 |
|
✗ |
STOP_PROFILER("SolverFDDP::solve"); |
| 103 |
|
✗ |
return false; |
| 104 |
|
|
} |
| 105 |
|
|
} |
| 106 |
|
✗ |
stoppingCriteria(); |
| 107 |
|
|
|
| 108 |
|
✗ |
const std::size_t n_callbacks = callbacks_.size(); |
| 109 |
|
✗ |
for (std::size_t c = 0; c < n_callbacks; ++c) { |
| 110 |
|
✗ |
CallbackAbstract& callback = *callbacks_[c]; |
| 111 |
|
✗ |
callback(*this); |
| 112 |
|
|
} |
| 113 |
|
|
|
| 114 |
|
✗ |
if (was_feasible_ && stop_ < th_stop_) { |
| 115 |
|
✗ |
STOP_PROFILER("SolverFDDP::solve"); |
| 116 |
|
✗ |
return true; |
| 117 |
|
|
} |
| 118 |
|
|
} |
| 119 |
|
✗ |
STOP_PROFILER("SolverFDDP::solve"); |
| 120 |
|
✗ |
return false; |
| 121 |
|
|
} |
| 122 |
|
|
|
| 123 |
|
✗ |
const Eigen::Vector2d& SolverFDDP::expectedImprovement() { |
| 124 |
|
✗ |
dv_ = 0; |
| 125 |
|
✗ |
const std::size_t T = this->problem_->get_T(); |
| 126 |
|
✗ |
if (!is_feasible_) { |
| 127 |
|
|
// NB: The dimension of vectors xs_try_ and xs_ are T+1, whereas the |
| 128 |
|
|
// dimension of dx_ is T. Here, we are re-using the final element of dx_ for |
| 129 |
|
|
// the computation of the difference at the terminal node. Using the access |
| 130 |
|
|
// iterator back() this re-use of the final element is fine. Cf. the |
| 131 |
|
|
// discussion at https://github.com/loco-3d/crocoddyl/issues/1022 |
| 132 |
|
✗ |
problem_->get_terminalModel()->get_state()->diff(xs_try_.back(), xs_.back(), |
| 133 |
|
✗ |
dx_.back()); |
| 134 |
|
✗ |
fTVxx_p_.noalias() = Vxx_.back() * dx_.back(); |
| 135 |
|
✗ |
dv_ -= fs_.back().dot(fTVxx_p_); |
| 136 |
|
|
const std::vector<std::shared_ptr<ActionModelAbstract> >& models = |
| 137 |
|
✗ |
problem_->get_runningModels(); |
| 138 |
|
|
|
| 139 |
|
✗ |
for (std::size_t t = 0; t < T; ++t) { |
| 140 |
|
✗ |
models[t]->get_state()->diff(xs_try_[t], xs_[t], dx_[t]); |
| 141 |
|
✗ |
fTVxx_p_.noalias() = Vxx_[t] * dx_[t]; |
| 142 |
|
✗ |
dv_ -= fs_[t].dot(fTVxx_p_); |
| 143 |
|
|
} |
| 144 |
|
|
} |
| 145 |
|
✗ |
d_[0] = dg_ + dv_; |
| 146 |
|
✗ |
d_[1] = dq_ - 2 * dv_; |
| 147 |
|
✗ |
return d_; |
| 148 |
|
|
} |
| 149 |
|
|
|
| 150 |
|
✗ |
void SolverFDDP::updateExpectedImprovement() { |
| 151 |
|
✗ |
dg_ = 0; |
| 152 |
|
✗ |
dq_ = 0; |
| 153 |
|
✗ |
const std::size_t T = this->problem_->get_T(); |
| 154 |
|
✗ |
if (!is_feasible_) { |
| 155 |
|
✗ |
dg_ -= Vx_.back().dot(fs_.back()); |
| 156 |
|
✗ |
fTVxx_p_.noalias() = Vxx_.back() * fs_.back(); |
| 157 |
|
✗ |
dq_ += fs_.back().dot(fTVxx_p_); |
| 158 |
|
|
} |
| 159 |
|
|
const std::vector<std::shared_ptr<ActionModelAbstract> >& models = |
| 160 |
|
✗ |
problem_->get_runningModels(); |
| 161 |
|
✗ |
for (std::size_t t = 0; t < T; ++t) { |
| 162 |
|
✗ |
const std::size_t nu = models[t]->get_nu(); |
| 163 |
|
✗ |
if (nu != 0) { |
| 164 |
|
✗ |
dg_ += Qu_[t].dot(k_[t]); |
| 165 |
|
✗ |
dq_ -= k_[t].dot(Quuk_[t]); |
| 166 |
|
|
} |
| 167 |
|
✗ |
if (!is_feasible_) { |
| 168 |
|
✗ |
dg_ -= Vx_[t].dot(fs_[t]); |
| 169 |
|
✗ |
fTVxx_p_.noalias() = Vxx_[t] * fs_[t]; |
| 170 |
|
✗ |
dq_ += fs_[t].dot(fTVxx_p_); |
| 171 |
|
|
} |
| 172 |
|
|
} |
| 173 |
|
|
} |
| 174 |
|
|
|
| 175 |
|
✗ |
void SolverFDDP::forwardPass(const double steplength) { |
| 176 |
|
✗ |
if (steplength > 1. || steplength < 0.) { |
| 177 |
|
✗ |
throw_pretty("Invalid argument: " |
| 178 |
|
|
<< "invalid step length, value is between 0. to 1."); |
| 179 |
|
|
} |
| 180 |
|
✗ |
START_PROFILER("SolverFDDP::forwardPass"); |
| 181 |
|
✗ |
cost_try_ = 0.; |
| 182 |
|
✗ |
xnext_ = problem_->get_x0(); |
| 183 |
|
✗ |
const std::size_t T = problem_->get_T(); |
| 184 |
|
|
const std::vector<std::shared_ptr<ActionModelAbstract> >& models = |
| 185 |
|
✗ |
problem_->get_runningModels(); |
| 186 |
|
|
const std::vector<std::shared_ptr<ActionDataAbstract> >& datas = |
| 187 |
|
✗ |
problem_->get_runningDatas(); |
| 188 |
|
✗ |
if ((is_feasible_) || (steplength == 1)) { |
| 189 |
|
✗ |
for (std::size_t t = 0; t < T; ++t) { |
| 190 |
|
✗ |
const std::shared_ptr<ActionModelAbstract>& m = models[t]; |
| 191 |
|
✗ |
const std::shared_ptr<ActionDataAbstract>& d = datas[t]; |
| 192 |
|
✗ |
const std::size_t nu = m->get_nu(); |
| 193 |
|
|
|
| 194 |
|
✗ |
xs_try_[t] = xnext_; |
| 195 |
|
✗ |
m->get_state()->diff(xs_[t], xs_try_[t], dx_[t]); |
| 196 |
|
✗ |
if (nu != 0) { |
| 197 |
|
✗ |
us_try_[t].noalias() = us_[t] - k_[t] * steplength - K_[t] * dx_[t]; |
| 198 |
|
✗ |
m->calc(d, xs_try_[t], us_try_[t]); |
| 199 |
|
|
} else { |
| 200 |
|
✗ |
m->calc(d, xs_try_[t]); |
| 201 |
|
|
} |
| 202 |
|
✗ |
xnext_ = d->xnext; |
| 203 |
|
✗ |
cost_try_ += d->cost; |
| 204 |
|
|
|
| 205 |
|
✗ |
if (raiseIfNaN(cost_try_)) { |
| 206 |
|
✗ |
STOP_PROFILER("SolverFDDP::forwardPass"); |
| 207 |
|
✗ |
throw_pretty("forward_error"); |
| 208 |
|
|
} |
| 209 |
|
✗ |
if (raiseIfNaN(xnext_.lpNorm<Eigen::Infinity>())) { |
| 210 |
|
✗ |
STOP_PROFILER("SolverFDDP::forwardPass"); |
| 211 |
|
✗ |
throw_pretty("forward_error"); |
| 212 |
|
|
} |
| 213 |
|
|
} |
| 214 |
|
|
|
| 215 |
|
|
const std::shared_ptr<ActionModelAbstract>& m = |
| 216 |
|
✗ |
problem_->get_terminalModel(); |
| 217 |
|
✗ |
const std::shared_ptr<ActionDataAbstract>& d = problem_->get_terminalData(); |
| 218 |
|
✗ |
xs_try_.back() = xnext_; |
| 219 |
|
✗ |
m->calc(d, xs_try_.back()); |
| 220 |
|
✗ |
cost_try_ += d->cost; |
| 221 |
|
|
|
| 222 |
|
✗ |
if (raiseIfNaN(cost_try_)) { |
| 223 |
|
✗ |
STOP_PROFILER("SolverFDDP::forwardPass"); |
| 224 |
|
✗ |
throw_pretty("forward_error"); |
| 225 |
|
|
} |
| 226 |
|
✗ |
} else { |
| 227 |
|
✗ |
for (std::size_t t = 0; t < T; ++t) { |
| 228 |
|
✗ |
const std::shared_ptr<ActionModelAbstract>& m = models[t]; |
| 229 |
|
✗ |
const std::shared_ptr<ActionDataAbstract>& d = datas[t]; |
| 230 |
|
✗ |
const std::size_t nu = m->get_nu(); |
| 231 |
|
✗ |
m->get_state()->integrate(xnext_, fs_[t] * (steplength - 1), xs_try_[t]); |
| 232 |
|
✗ |
m->get_state()->diff(xs_[t], xs_try_[t], dx_[t]); |
| 233 |
|
✗ |
if (nu != 0) { |
| 234 |
|
✗ |
us_try_[t].noalias() = us_[t] - k_[t] * steplength - K_[t] * dx_[t]; |
| 235 |
|
✗ |
m->calc(d, xs_try_[t], us_try_[t]); |
| 236 |
|
|
} else { |
| 237 |
|
✗ |
m->calc(d, xs_try_[t]); |
| 238 |
|
|
} |
| 239 |
|
✗ |
xnext_ = d->xnext; |
| 240 |
|
✗ |
cost_try_ += d->cost; |
| 241 |
|
|
|
| 242 |
|
✗ |
if (raiseIfNaN(cost_try_)) { |
| 243 |
|
✗ |
STOP_PROFILER("SolverFDDP::forwardPass"); |
| 244 |
|
✗ |
throw_pretty("forward_error"); |
| 245 |
|
|
} |
| 246 |
|
✗ |
if (raiseIfNaN(xnext_.lpNorm<Eigen::Infinity>())) { |
| 247 |
|
✗ |
STOP_PROFILER("SolverFDDP::forwardPass"); |
| 248 |
|
✗ |
throw_pretty("forward_error"); |
| 249 |
|
|
} |
| 250 |
|
|
} |
| 251 |
|
|
|
| 252 |
|
|
const std::shared_ptr<ActionModelAbstract>& m = |
| 253 |
|
✗ |
problem_->get_terminalModel(); |
| 254 |
|
✗ |
const std::shared_ptr<ActionDataAbstract>& d = problem_->get_terminalData(); |
| 255 |
|
✗ |
m->get_state()->integrate(xnext_, fs_.back() * (steplength - 1), |
| 256 |
|
✗ |
xs_try_.back()); |
| 257 |
|
✗ |
m->calc(d, xs_try_.back()); |
| 258 |
|
✗ |
cost_try_ += d->cost; |
| 259 |
|
|
|
| 260 |
|
✗ |
if (raiseIfNaN(cost_try_)) { |
| 261 |
|
✗ |
STOP_PROFILER("SolverFDDP::forwardPass"); |
| 262 |
|
✗ |
throw_pretty("forward_error"); |
| 263 |
|
|
} |
| 264 |
|
|
} |
| 265 |
|
✗ |
STOP_PROFILER("SolverFDDP::forwardPass"); |
| 266 |
|
|
} |
| 267 |
|
|
|
| 268 |
|
✗ |
double SolverFDDP::get_th_acceptnegstep() const { return th_acceptnegstep_; } |
| 269 |
|
|
|
| 270 |
|
✗ |
void SolverFDDP::set_th_acceptnegstep(const double th_acceptnegstep) { |
| 271 |
|
✗ |
if (0. > th_acceptnegstep) { |
| 272 |
|
✗ |
throw_pretty( |
| 273 |
|
|
"Invalid argument: " << "th_acceptnegstep value has to be positive."); |
| 274 |
|
|
} |
| 275 |
|
✗ |
th_acceptnegstep_ = th_acceptnegstep; |
| 276 |
|
|
} |
| 277 |
|
|
|
| 278 |
|
|
} // namespace crocoddyl |
| 279 |
|
|
|