Line |
Branch |
Exec |
Source |
1 |
|
|
// |
2 |
|
|
// Copyright (c) 2022-2024 INRIA |
3 |
|
|
// |
4 |
|
|
|
5 |
|
|
#ifndef __pinocchio_algorithm_admm_solver_hxx__ |
6 |
|
|
#define __pinocchio_algorithm_admm_solver_hxx__ |
7 |
|
|
|
8 |
|
|
#include <limits> |
9 |
|
|
#include <iomanip> |
10 |
|
|
#include <iostream> |
11 |
|
|
|
12 |
|
|
#include "pinocchio/algorithm/contact-solver-utils.hpp" |
13 |
|
|
#include "pinocchio/algorithm/constraints/coulomb-friction-cone.hpp" |
14 |
|
|
|
15 |
|
|
namespace pinocchio |
16 |
|
|
{ |
17 |
|
|
|
18 |
|
|
template<typename _Scalar> |
19 |
|
|
template< |
20 |
|
|
typename DelassusDerived, |
21 |
|
|
typename VectorLike, |
22 |
|
|
typename ConstraintAllocator, |
23 |
|
|
typename VectorLikeR> |
24 |
|
✗ |
bool ADMMContactSolverTpl<_Scalar>::solve( |
25 |
|
|
DelassusOperatorBase<DelassusDerived> & _delassus, |
26 |
|
|
const Eigen::MatrixBase<VectorLike> & g, |
27 |
|
|
const std::vector<CoulombFrictionConeTpl<Scalar>, ConstraintAllocator> & cones, |
28 |
|
|
const Eigen::MatrixBase<VectorLikeR> & R, |
29 |
|
|
const boost::optional<ConstRefVectorXs> primal_guess, |
30 |
|
|
const boost::optional<ConstRefVectorXs> dual_guess, |
31 |
|
|
bool compute_largest_eigen_values, |
32 |
|
|
bool stat_record) |
33 |
|
|
|
34 |
|
|
{ |
35 |
|
|
using namespace internal; |
36 |
|
|
|
37 |
|
✗ |
DelassusDerived & delassus = _delassus.derived(); |
38 |
|
|
|
39 |
|
✗ |
const Scalar mu_R = R.minCoeff(); |
40 |
|
✗ |
PINOCCHIO_CHECK_INPUT_ARGUMENT(tau <= Scalar(1) && tau > Scalar(0), "tau should lie in ]0,1]."); |
41 |
|
✗ |
PINOCCHIO_CHECK_INPUT_ARGUMENT(mu_prox >= 0, "mu_prox should be positive."); |
42 |
|
✗ |
PINOCCHIO_CHECK_INPUT_ARGUMENT(mu_R >= Scalar(0), "R should be a positive vector."); |
43 |
|
✗ |
PINOCCHIO_CHECK_ARGUMENT_SIZE(R.size(), problem_size); |
44 |
|
|
// PINOCCHIO_CHECK_INPUT_ARGUMENT(math::max(R.maxCoeff(),mu_prox) >= 0,"mu_prox and mu_R |
45 |
|
|
// cannot be both equal to zero."); |
46 |
|
|
|
47 |
|
✗ |
if (compute_largest_eigen_values) |
48 |
|
|
{ |
49 |
|
|
// const Scalar L = delassus.computeLargestEigenValue(20); // Largest eigen_value |
50 |
|
|
// estimate. |
51 |
|
✗ |
power_iteration_algo.run(delassus); |
52 |
|
|
} |
53 |
|
✗ |
const Scalar L = power_iteration_algo.largest_eigen_value; |
54 |
|
|
// const Scalar L = delassus.computeLargestEigenValue(20); |
55 |
|
✗ |
const Scalar m = mu_prox + mu_R; |
56 |
|
✗ |
const Scalar cond = L / m; |
57 |
|
✗ |
const Scalar rho_increment = std::pow(cond, rho_power_factor); |
58 |
|
|
|
59 |
|
|
Scalar complementarity, |
60 |
|
|
proximal_metric, // proximal metric between two successive iterates. |
61 |
|
|
primal_feasibility, dual_feasibility_ncp, dual_feasibility; |
62 |
|
|
|
63 |
|
|
// std::cout << std::setprecision(12); |
64 |
|
|
|
65 |
|
|
Scalar rho; |
66 |
|
✗ |
rho = computeRho(L, m, rho_power); |
67 |
|
|
// if(!is_initialized) |
68 |
|
|
// { |
69 |
|
|
// rho = computeRho(L,m,rho_power); |
70 |
|
|
// } |
71 |
|
|
// else |
72 |
|
|
// { |
73 |
|
|
// rho = this->rho; |
74 |
|
|
// } |
75 |
|
|
// rho = computeRho(L,m,rho_power); |
76 |
|
|
|
77 |
|
|
// std::cout << "L: " << L << std::endl; |
78 |
|
|
// std::cout << "m: " << m << std::endl; |
79 |
|
|
// std::cout << "prox_value: " << prox_value << std::endl; |
80 |
|
|
|
81 |
|
|
PINOCCHIO_EIGEN_MALLOC_NOT_ALLOWED(); |
82 |
|
|
|
83 |
|
|
// Update the cholesky decomposition |
84 |
|
✗ |
Scalar prox_value = mu_prox + tau * rho; |
85 |
|
✗ |
rhs = R + VectorXs::Constant(this->problem_size, prox_value); |
86 |
|
✗ |
delassus.updateDamping(rhs); |
87 |
|
✗ |
cholesky_update_count = 1; |
88 |
|
|
|
89 |
|
|
// Initial update of the variables |
90 |
|
|
// Init x |
91 |
|
✗ |
if (primal_guess) |
92 |
|
|
{ |
93 |
|
✗ |
x_ = primal_guess.get(); |
94 |
|
✗ |
PINOCCHIO_CHECK_ARGUMENT_SIZE(x_.size(), problem_size); |
95 |
|
|
} |
96 |
|
✗ |
else if (!is_initialized) |
97 |
|
|
{ |
98 |
|
✗ |
x_.setZero(); |
99 |
|
|
} |
100 |
|
|
else |
101 |
|
|
{ |
102 |
|
✗ |
x_ = y_; // takes the current value stored in the solver |
103 |
|
|
} |
104 |
|
|
|
105 |
|
|
// Init y |
106 |
|
✗ |
computeConeProjection(cones, x_, y_); |
107 |
|
|
|
108 |
|
|
// Init z |
109 |
|
✗ |
if (dual_guess) |
110 |
|
|
{ |
111 |
|
✗ |
z_ = dual_guess.get(); |
112 |
|
✗ |
PINOCCHIO_CHECK_ARGUMENT_SIZE(z_.size(), problem_size); |
113 |
|
|
} |
114 |
|
✗ |
else if (!is_initialized) |
115 |
|
|
{ |
116 |
|
✗ |
delassus.applyOnTheRight(y_, z_); // z = G * y |
117 |
|
✗ |
z_.noalias() += -prox_value * y_ + g; |
118 |
|
✗ |
computeComplementarityShift(cones, z_, s_); |
119 |
|
✗ |
z_ += s_; // Add De Saxé shift |
120 |
|
✗ |
computeDualConeProjection(cones, z_, z_); |
121 |
|
|
} |
122 |
|
|
else |
123 |
|
|
{ |
124 |
|
|
// Keep z from the previous iteration |
125 |
|
|
} |
126 |
|
|
|
127 |
|
|
// std::cout << "x_: " << x_.transpose() << std::endl; |
128 |
|
|
// std::cout << "y_: " << y_.transpose() << std::endl; |
129 |
|
|
// std::cout << "z_: " << z_.transpose() << std::endl; |
130 |
|
|
|
131 |
|
✗ |
if (stat_record) |
132 |
|
|
{ |
133 |
|
✗ |
stats.reset(); |
134 |
|
|
|
135 |
|
|
// Compute initial problem primal and dual feasibility |
136 |
|
✗ |
primal_feasibility_vector = x_ - y_; |
137 |
|
✗ |
primal_feasibility = primal_feasibility_vector.template lpNorm<Eigen::Infinity>(); |
138 |
|
|
} |
139 |
|
|
|
140 |
|
✗ |
is_initialized = true; |
141 |
|
|
|
142 |
|
|
// End of Initialization phase |
143 |
|
|
|
144 |
|
✗ |
bool abs_prec_reached = false, rel_prec_reached = false; |
145 |
|
|
|
146 |
|
✗ |
Scalar y_previous_norm_inf = y_.template lpNorm<Eigen::Infinity>(); |
147 |
|
✗ |
int it = 1; |
148 |
|
|
// Scalar res = 0; |
149 |
|
|
#ifdef PINOCCHIO_WITH_HPP_FCL |
150 |
|
✗ |
timer.start(); |
151 |
|
|
#endif // PINOCCHIO_WITH_HPP_FCL |
152 |
|
✗ |
for (; it <= Base::max_it; ++it) |
153 |
|
|
{ |
154 |
|
|
// std::cout << "---" << std::endl; |
155 |
|
|
// std::cout << "it: " << it << std::endl; |
156 |
|
|
// std::cout << "tau*rho: " << tau*rho << std::endl; |
157 |
|
|
|
158 |
|
✗ |
x_previous = x_; |
159 |
|
✗ |
y_previous = y_; |
160 |
|
✗ |
z_previous = z_; |
161 |
|
✗ |
complementarity = Scalar(0); |
162 |
|
|
|
163 |
|
|
// s-update |
164 |
|
✗ |
computeComplementarityShift(cones, z_, s_); |
165 |
|
|
|
166 |
|
|
// std::cout << "s_: " << s_.transpose() << std::endl; |
167 |
|
|
|
168 |
|
|
// std::cout << "x_: " << x_.transpose() << std::endl; |
169 |
|
|
|
170 |
|
|
// z-update |
171 |
|
|
// const Scalar alpha = 1.; |
172 |
|
|
// z_ -= (tau*rho) * (x_ - y_); |
173 |
|
|
// std::cout << "intermediate z_: " << z_.transpose() << std::endl; |
174 |
|
|
|
175 |
|
|
// x-update |
176 |
|
✗ |
rhs = -(g + s_ - (rho * tau) * y_ - mu_prox * x_ - z_); |
177 |
|
✗ |
const VectorXs rhs_copy = rhs; |
178 |
|
|
// x_ = rhs; |
179 |
|
✗ |
delassus.solveInPlace(rhs); |
180 |
|
|
// VectorXs tmp = delassus * rhs - rhs_copy; |
181 |
|
|
// res = math::max(res,tmp.template lpNorm<Eigen::Infinity>()); |
182 |
|
|
// std::cout << "residual = " << (delassus * rhs - x_).template lpNorm<Eigen::Infinity>() |
183 |
|
|
// << std::endl; |
184 |
|
✗ |
x_ = rhs; |
185 |
|
|
|
186 |
|
|
// y-update |
187 |
|
|
// rhs *= alpha; |
188 |
|
|
// rhs += (1-alpha)*y_previous; |
189 |
|
✗ |
rhs = x_; |
190 |
|
✗ |
rhs -= z_ / (tau * rho); |
191 |
|
✗ |
computeConeProjection(cones, rhs, y_); |
192 |
|
|
// std::cout << "y_: " << y_.transpose() << std::endl; |
193 |
|
|
|
194 |
|
|
// z-update |
195 |
|
✗ |
z_ -= (tau * rho) * (x_ - y_); |
196 |
|
|
// const Scalar gamma = Scalar(it) / Scalar(it + 300); |
197 |
|
|
|
198 |
|
|
// z_ += gamma * (z_ - z_previous).eval(); |
199 |
|
|
// x_ += gamma * (x_ - x_previous).eval(); |
200 |
|
|
// computeConeProjection(cones, y_, y_); |
201 |
|
|
|
202 |
|
|
// z_ -= (tau*rho) * (x_ * alpha + (1-alpha)*y_previous - y_); |
203 |
|
|
// std::cout << "z_: " << z_.transpose() << std::endl; |
204 |
|
|
// computeDualConeProjection(cones, z_, z_); |
205 |
|
|
|
206 |
|
|
// check termination criteria |
207 |
|
✗ |
primal_feasibility_vector = x_ - y_; |
208 |
|
|
// delassus.applyOnTheRight(x_,dual_feasibility_vector); |
209 |
|
|
// dual_feasibility_vector.noalias() += g + s_ - prox_value * x_ - z_; |
210 |
|
|
|
211 |
|
|
{ |
212 |
|
✗ |
VectorXs & dy = rhs; |
213 |
|
✗ |
dy = y_ - y_previous; |
214 |
|
✗ |
proximal_metric = dy.template lpNorm<Eigen::Infinity>(); |
215 |
|
✗ |
dual_feasibility_vector.noalias() = (tau * rho) * dy; |
216 |
|
|
} |
217 |
|
|
|
218 |
|
|
{ |
219 |
|
✗ |
VectorXs & dx = rhs; |
220 |
|
✗ |
dx = x_ - x_previous; |
221 |
|
✗ |
dual_feasibility_vector.noalias() += mu_prox * dx; |
222 |
|
|
} |
223 |
|
|
|
224 |
|
|
// delassus.applyOnTheRight(x_,dual_feasibility_vector); |
225 |
|
|
// dual_feasibility_vector.noalias() += g; |
226 |
|
|
// computeComplementarityShift(cones, z_, s_); |
227 |
|
|
// dual_feasibility_vector.noalias() += s_ - prox_value * x_ - z_; |
228 |
|
|
|
229 |
|
✗ |
primal_feasibility = primal_feasibility_vector.template lpNorm<Eigen::Infinity>(); |
230 |
|
✗ |
dual_feasibility = dual_feasibility_vector.template lpNorm<Eigen::Infinity>(); |
231 |
|
✗ |
complementarity = computeConicComplementarity(cones, z_, y_); |
232 |
|
|
// complementarity = z_.dot(y_)/cones.size(); |
233 |
|
|
|
234 |
|
✗ |
if (stat_record) |
235 |
|
|
{ |
236 |
|
✗ |
VectorXs tmp(rhs); |
237 |
|
✗ |
delassus.applyOnTheRight(y_, rhs); |
238 |
|
✗ |
rhs.noalias() += g - prox_value * y_; |
239 |
|
✗ |
computeComplementarityShift(cones, rhs, tmp); |
240 |
|
✗ |
rhs.noalias() += tmp; |
241 |
|
|
|
242 |
|
✗ |
internal::computeDualConeProjection(cones, rhs, tmp); |
243 |
|
✗ |
tmp -= rhs; |
244 |
|
|
|
245 |
|
✗ |
dual_feasibility_ncp = tmp.template lpNorm<Eigen::Infinity>(); |
246 |
|
|
|
247 |
|
✗ |
stats.primal_feasibility.push_back(primal_feasibility); |
248 |
|
✗ |
stats.dual_feasibility.push_back(dual_feasibility); |
249 |
|
✗ |
stats.dual_feasibility_ncp.push_back(dual_feasibility_ncp); |
250 |
|
✗ |
stats.complementarity.push_back(complementarity); |
251 |
|
✗ |
stats.rho.push_back(rho); |
252 |
|
|
} |
253 |
|
|
|
254 |
|
|
// std::cout << "primal_feasibility: " << primal_feasibility << std::endl; |
255 |
|
|
// std::cout << "dual_feasibility: " << dual_feasibility << std::endl; |
256 |
|
|
// std::cout << "complementarity: " << complementarity << std::endl; |
257 |
|
|
|
258 |
|
|
// Checking stopping residual |
259 |
|
✗ |
if ( |
260 |
|
✗ |
check_expression_if_real<Scalar, false>(complementarity <= this->absolute_precision) |
261 |
|
✗ |
&& check_expression_if_real<Scalar, false>(dual_feasibility <= this->absolute_precision) |
262 |
|
✗ |
&& check_expression_if_real<Scalar, false>(primal_feasibility <= this->absolute_precision)) |
263 |
|
✗ |
abs_prec_reached = true; |
264 |
|
|
else |
265 |
|
✗ |
abs_prec_reached = false; |
266 |
|
|
|
267 |
|
✗ |
const Scalar y_norm_inf = y_.template lpNorm<Eigen::Infinity>(); |
268 |
|
✗ |
if (check_expression_if_real<Scalar, false>( |
269 |
|
|
proximal_metric |
270 |
|
✗ |
<= this->relative_precision * math::max(y_norm_inf, y_previous_norm_inf))) |
271 |
|
✗ |
rel_prec_reached = true; |
272 |
|
|
else |
273 |
|
✗ |
rel_prec_reached = false; |
274 |
|
|
|
275 |
|
|
// if(abs_prec_reached || rel_prec_reached) |
276 |
|
✗ |
if (abs_prec_reached) |
277 |
|
✗ |
break; |
278 |
|
|
|
279 |
|
|
// Account for potential update of rho |
280 |
|
✗ |
bool update_delassus_factorization = false; |
281 |
|
✗ |
if (primal_feasibility > ratio_primal_dual * dual_feasibility) |
282 |
|
|
{ |
283 |
|
✗ |
rho *= rho_increment; |
284 |
|
|
// rho *= math::pow(cond,rho_power_factor); |
285 |
|
|
// rho_power += rho_power_factor; |
286 |
|
✗ |
update_delassus_factorization = true; |
287 |
|
|
} |
288 |
|
✗ |
else if (dual_feasibility > ratio_primal_dual * primal_feasibility) |
289 |
|
|
{ |
290 |
|
✗ |
rho /= rho_increment; |
291 |
|
|
// rho *= math::pow(cond,-rho_power_factor); |
292 |
|
|
// rho_power -= rho_power_factor; |
293 |
|
✗ |
update_delassus_factorization = true; |
294 |
|
|
} |
295 |
|
|
|
296 |
|
✗ |
if (update_delassus_factorization) |
297 |
|
|
{ |
298 |
|
✗ |
prox_value = mu_prox + tau * rho; |
299 |
|
✗ |
rhs = R + VectorXs::Constant(this->problem_size, prox_value); |
300 |
|
✗ |
delassus.updateDamping(rhs); |
301 |
|
✗ |
cholesky_update_count++; |
302 |
|
|
} |
303 |
|
|
|
304 |
|
✗ |
y_previous_norm_inf = y_norm_inf; |
305 |
|
|
// std::cout << "rho_power: " << rho_power << std::endl; |
306 |
|
|
// std::cout << "rho: " << rho << std::endl; |
307 |
|
|
// std::cout << "---" << std::endl; |
308 |
|
|
} |
309 |
|
|
|
310 |
|
|
PINOCCHIO_EIGEN_MALLOC_ALLOWED(); |
311 |
|
|
|
312 |
|
✗ |
this->absolute_residual = |
313 |
|
✗ |
math::max(primal_feasibility, math::max(complementarity, dual_feasibility)); |
314 |
|
✗ |
this->relative_residual = proximal_metric; |
315 |
|
✗ |
this->it = it; |
316 |
|
|
// std::cout << "max linalg res: " << res << std::endl; |
317 |
|
|
// y_sol.const_cast_derived() = y_; |
318 |
|
|
|
319 |
|
|
// Save values |
320 |
|
✗ |
this->rho_power = computeRhoPower(L, m, rho); |
321 |
|
✗ |
this->rho = rho; |
322 |
|
|
|
323 |
|
✗ |
if (stat_record) |
324 |
|
|
{ |
325 |
|
✗ |
stats.it = it; |
326 |
|
✗ |
stats.cholesky_update_count = cholesky_update_count; |
327 |
|
|
} |
328 |
|
|
|
329 |
|
|
#ifdef PINOCCHIO_WITH_HPP_FCL |
330 |
|
✗ |
timer.stop(); |
331 |
|
|
#endif // PINOCCHIO_WITH_HPP_FCL |
332 |
|
|
|
333 |
|
|
// if(abs_prec_reached || rel_prec_reached) |
334 |
|
✗ |
if (abs_prec_reached) |
335 |
|
✗ |
return true; |
336 |
|
|
|
337 |
|
✗ |
return false; |
338 |
|
|
} |
339 |
|
|
} // namespace pinocchio |
340 |
|
|
|
341 |
|
|
#endif // ifndef __pinocchio_algorithm_admm_solver_hxx__ |
342 |
|
|
|