| 1 | /* ---------------------------------------------------------------------------- |
| 2 | |
| 3 | * GTSAM Copyright 2010, Georgia Tech Research Corporation, |
| 4 | * Atlanta, Georgia 30332-0415 |
| 5 | * All Rights Reserved |
| 6 | * Authors: Frank Dellaert, et al. (see THANKS for the full author list) |
| 7 | |
| 8 | * See LICENSE for the license information |
| 9 | |
| 10 | * -------------------------------------------------------------------------- */ |
| 11 | |
| 12 | /** |
| 13 | * @file ActiveSetSolver-inl.h |
| 14 | * @brief Implmentation of ActiveSetSolver. |
| 15 | * @author Ivan Dario Jimenez |
| 16 | * @author Duy Nguyen Ta |
| 17 | * @date 2/11/16 |
| 18 | */ |
| 19 | |
| 20 | #pragma once |
| 21 | |
| 22 | #include <gtsam_unstable/linear/InfeasibleInitialValues.h> |
| 23 | |
| 24 | /******************************************************************************/ |
| 25 | // Convenient macros to reduce syntactic noise. undef later. |
| 26 | #define Template template <class PROBLEM, class POLICY, class INITSOLVER> |
| 27 | #define This ActiveSetSolver<PROBLEM, POLICY, INITSOLVER> |
| 28 | |
| 29 | /******************************************************************************/ |
| 30 | |
| 31 | namespace gtsam { |
| 32 | |
| 33 | /* We have to make sure the new solution with alpha satisfies all INACTIVE inequality constraints |
| 34 | * If some inactive inequality constraints complain about the full step (alpha = 1), |
| 35 | * we have to adjust alpha to stay within the inequality constraints' feasible regions. |
| 36 | * |
| 37 | * For each inactive inequality j: |
| 38 | * - We already have: aj'*xk - bj <= 0, since xk satisfies all inequality constraints |
| 39 | * - We want: aj'*(xk + alpha*p) - bj <= 0 |
| 40 | * - If aj'*p <= 0, we have: aj'*(xk + alpha*p) <= aj'*xk <= bj, for all alpha>0 |
| 41 | * it's good! |
| 42 | * - We only care when aj'*p > 0. In this case, we need to choose alpha so that |
| 43 | * aj'*xk + alpha*aj'*p - bj <= 0 --> alpha <= (bj - aj'*xk) / (aj'*p) |
| 44 | * We want to step as far as possible, so we should choose alpha = (bj - aj'*xk) / (aj'*p) |
| 45 | * |
| 46 | * We want the minimum of all those alphas among all inactive inequality. |
| 47 | */ |
| 48 | Template std::tuple<double, int> This::computeStepSize( |
| 49 | const InequalityFactorGraph& workingSet, const VectorValues& xk, |
| 50 | const VectorValues& p, const double& maxAlpha) const { |
| 51 | double minAlpha = maxAlpha; |
| 52 | int closestFactorIx = -1; |
| 53 | for (size_t factorIx = 0; factorIx < workingSet.size(); ++factorIx) { |
| 54 | const LinearInequality::shared_ptr& factor = workingSet.at(i: factorIx); |
| 55 | double b = factor->getb()[0]; |
| 56 | // only check inactive factors |
| 57 | if (!factor->active()) { |
| 58 | // Compute a'*p |
| 59 | double aTp = factor->dotProductRow(p); |
| 60 | |
| 61 | // Check if a'*p >0. Don't care if it's not. |
| 62 | if (aTp <= 0) |
| 63 | continue; |
| 64 | |
| 65 | // Compute a'*xk |
| 66 | double aTx = factor->dotProductRow(p: xk); |
| 67 | |
| 68 | // alpha = (b - a'*xk) / (a'*p) |
| 69 | double alpha = (b - aTx) / aTp; |
| 70 | // We want the minimum of all those max alphas |
| 71 | if (alpha < minAlpha) { |
| 72 | closestFactorIx = factorIx; |
| 73 | minAlpha = alpha; |
| 74 | } |
| 75 | } |
| 76 | } |
| 77 | return std::make_tuple(args&: minAlpha, args&: closestFactorIx); |
| 78 | } |
| 79 | |
| 80 | /******************************************************************************/ |
| 81 | /* |
| 82 | * The goal of this function is to find currently active inequality constraints |
| 83 | * that violate the condition to be active. The one that violates the condition |
| 84 | * the most will be removed from the active set. See Nocedal06book, pg 469-471 |
| 85 | * |
| 86 | * Find the BAD active inequality that pulls x strongest to the wrong direction |
| 87 | * of its constraint (i.e. it is pulling towards >0, while its feasible region is <=0) |
| 88 | * |
| 89 | * For active inequality constraints (those that are enforced as equality constraints |
| 90 | * in the current working set), we want lambda < 0. |
| 91 | * This is because: |
| 92 | * - From the Lagrangian L = f - lambda*c, we know that the constraint force |
| 93 | * is (lambda * \grad c) = \grad f. Intuitively, to keep the solution x stay |
| 94 | * on the constraint surface, the constraint force has to balance out with |
| 95 | * other unconstrained forces that are pulling x towards the unconstrained |
| 96 | * minimum point. The other unconstrained forces are pulling x toward (-\grad f), |
| 97 | * hence the constraint force has to be exactly \grad f, so that the total |
| 98 | * force is 0. |
| 99 | * - We also know that at the constraint surface c(x)=0, \grad c points towards + (>= 0), |
| 100 | * while we are solving for - (<=0) constraint. |
| 101 | * - We want the constraint force (lambda * \grad c) to pull x towards the - (<=0) direction |
| 102 | * i.e., the opposite direction of \grad c where the inequality constraint <=0 is satisfied. |
| 103 | * That means we want lambda < 0. |
| 104 | * - This is because when the constrained force pulls x towards the infeasible region (+), |
| 105 | * the unconstrained force is pulling x towards the opposite direction into |
| 106 | * the feasible region (again because the total force has to be 0 to make x stay still) |
| 107 | * So we can drop this constraint to have a lower error but feasible solution. |
| 108 | * |
| 109 | * In short, active inequality constraints with lambda > 0 are BAD, because they |
| 110 | * violate the condition to be active. |
| 111 | * |
| 112 | * And we want to remove the worst one with the largest lambda from the active set. |
| 113 | * |
| 114 | */ |
| 115 | Template int This::identifyLeavingConstraint( |
| 116 | const InequalityFactorGraph& workingSet, |
| 117 | const VectorValues& lambdas) const { |
| 118 | int worstFactorIx = -1; |
| 119 | // preset the maxLambda to 0.0: if lambda is <= 0.0, the constraint is either |
| 120 | // inactive or a good inequality constraint, so we don't care! |
| 121 | double maxLambda = 0.0; |
| 122 | for (size_t factorIx = 0; factorIx < workingSet.size(); ++factorIx) { |
| 123 | const LinearInequality::shared_ptr& factor = workingSet.at(i: factorIx); |
| 124 | if (factor->active()) { |
| 125 | double lambda = lambdas.at(j: factor->dualKey())[0]; |
| 126 | if (lambda > maxLambda) { |
| 127 | worstFactorIx = factorIx; |
| 128 | maxLambda = lambda; |
| 129 | } |
| 130 | } |
| 131 | } |
| 132 | return worstFactorIx; |
| 133 | } |
| 134 | |
| 135 | //****************************************************************************** |
| 136 | Template JacobianFactor::shared_ptr This::createDualFactor( |
| 137 | Key key, const InequalityFactorGraph& workingSet, |
| 138 | const VectorValues& delta) const { |
| 139 | // Transpose the A matrix of constrained factors to have the jacobian of the |
| 140 | // dual key |
| 141 | TermsContainer Aterms = collectDualJacobians<LinearEquality>( |
| 142 | key, problem_.equalities, equalityVariableIndex_); |
| 143 | TermsContainer AtermsInequalities = collectDualJacobians<LinearInequality>( |
| 144 | key, workingSet, inequalityVariableIndex_); |
| 145 | Aterms.insert(position: Aterms.end(), first: AtermsInequalities.begin(), |
| 146 | last: AtermsInequalities.end()); |
| 147 | |
| 148 | // Collect the gradients of unconstrained cost factors to the b vector |
| 149 | if (Aterms.size() > 0) { |
| 150 | Vector b = problem_.costGradient(key, delta); |
| 151 | // to compute the least-square approximation of dual variables |
| 152 | return std::make_shared<JacobianFactor>(args&: Aterms, args&: b); |
| 153 | } else { |
| 154 | return nullptr; |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | /******************************************************************************/ |
| 159 | /* This function will create a dual graph that solves for the |
| 160 | * lagrange multipliers for the current working set. |
| 161 | * You can use lagrange multipliers as a necessary condition for optimality. |
| 162 | * The factor graph that is being solved is f' = -lambda * g' |
| 163 | * where f is the optimized function and g is the function resulting from |
| 164 | * aggregating the working set. |
| 165 | * The lambdas give you information about the feasibility of a constraint. |
| 166 | * if lambda < 0 the constraint is Ok |
| 167 | * if lambda = 0 you are on the constraint |
| 168 | * if lambda > 0 you are violating the constraint. |
| 169 | */ |
| 170 | Template GaussianFactorGraph This::buildDualGraph( |
| 171 | const InequalityFactorGraph& workingSet, const VectorValues& delta) const { |
| 172 | GaussianFactorGraph dualGraph; |
| 173 | for (Key key : constrainedKeys_) { |
| 174 | // Each constrained key becomes a factor in the dual graph |
| 175 | auto dualFactor = createDualFactor(key, workingSet, delta); |
| 176 | if (dualFactor) dualGraph.push_back(dualFactor); |
| 177 | } |
| 178 | return dualGraph; |
| 179 | } |
| 180 | |
| 181 | //****************************************************************************** |
| 182 | Template GaussianFactorGraph |
| 183 | This::buildWorkingGraph(const InequalityFactorGraph& workingSet, |
| 184 | const VectorValues& xk) const { |
| 185 | GaussianFactorGraph workingGraph; |
| 186 | workingGraph.push_back(POLICY::buildCostFunction(problem_, xk)); |
| 187 | workingGraph.push_back(problem_.equalities); |
| 188 | for (const LinearInequality::shared_ptr& factor : workingSet) |
| 189 | if (factor->active()) workingGraph.push_back(factor); |
| 190 | return workingGraph; |
| 191 | } |
| 192 | |
| 193 | //****************************************************************************** |
| 194 | Template typename This::State This::iterate( |
| 195 | const typename This::State& state) const { |
| 196 | // Algorithm 16.3 from Nocedal06book. |
| 197 | // Solve with the current working set eqn 16.39, but solve for x not p |
| 198 | auto workingGraph = buildWorkingGraph(workingSet: state.workingSet, xk: state.values); |
| 199 | VectorValues newValues = workingGraph.optimize(); |
| 200 | // If we CAN'T move further |
| 201 | // if p_k = 0 is the original condition, modified by Duy to say that the state |
| 202 | // update is zero. |
| 203 | if (newValues.equals(x: state.values, tol: 1e-7)) { |
| 204 | // Compute lambda from the dual graph |
| 205 | auto dualGraph = buildDualGraph(workingSet: state.workingSet, delta: newValues); |
| 206 | VectorValues duals = dualGraph.optimize(); |
| 207 | int leavingFactor = identifyLeavingConstraint(workingSet: state.workingSet, lambdas: duals); |
| 208 | // If all inequality constraints are satisfied: We have the solution!! |
| 209 | if (leavingFactor < 0) { |
| 210 | return State(newValues, duals, state.workingSet, true, |
| 211 | state.iterations + 1); |
| 212 | } else { |
| 213 | // Inactivate the leaving constraint |
| 214 | InequalityFactorGraph newWorkingSet = state.workingSet; |
| 215 | newWorkingSet.at(i: leavingFactor)->inactivate(); |
| 216 | return State(newValues, duals, newWorkingSet, false, |
| 217 | state.iterations + 1); |
| 218 | } |
| 219 | } else { |
| 220 | // If we CAN make some progress, i.e. p_k != 0 |
| 221 | // Adapt stepsize if some inactive constraints complain about this move |
| 222 | VectorValues p = newValues - state.values; |
| 223 | const auto [alpha, factorIx] = // using 16.41 |
| 224 | computeStepSize(workingSet: state.workingSet, xk: state.values, p, maxAlpha: POLICY::maxAlpha); |
| 225 | // also add to the working set the one that complains the most |
| 226 | InequalityFactorGraph newWorkingSet = state.workingSet; |
| 227 | if (factorIx >= 0) |
| 228 | newWorkingSet.at(factorIx)->activate(); |
| 229 | // step! |
| 230 | newValues = state.values + alpha * p; |
| 231 | return State(newValues, state.duals, newWorkingSet, false, |
| 232 | state.iterations + 1); |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | //****************************************************************************** |
| 237 | Template InequalityFactorGraph This::identifyActiveConstraints( |
| 238 | const InequalityFactorGraph& inequalities, |
| 239 | const VectorValues& initialValues, const VectorValues& duals, |
| 240 | bool useWarmStart) const { |
| 241 | InequalityFactorGraph workingSet; |
| 242 | for (const LinearInequality::shared_ptr& factor : inequalities) { |
| 243 | LinearInequality::shared_ptr workingFactor(new LinearInequality(*factor)); |
| 244 | if (useWarmStart && duals.size() > 0) { |
| 245 | if (duals.exists(j: workingFactor->dualKey())) workingFactor->activate(); |
| 246 | else workingFactor->inactivate(); |
| 247 | } else { |
| 248 | double error = workingFactor->error(c: initialValues); |
| 249 | // Safety guard. This should not happen unless users provide a bad init |
| 250 | if (error > 0) throw InfeasibleInitialValues(); |
| 251 | if (std::abs(x: error) < 1e-7) |
| 252 | workingFactor->activate(); |
| 253 | else |
| 254 | workingFactor->inactivate(); |
| 255 | } |
| 256 | workingSet.push_back(factor: workingFactor); |
| 257 | } |
| 258 | return workingSet; |
| 259 | } |
| 260 | |
| 261 | //****************************************************************************** |
| 262 | Template std::pair<VectorValues, VectorValues> This::optimize( |
| 263 | const VectorValues& initialValues, const VectorValues& duals, |
| 264 | bool useWarmStart) const { |
| 265 | // Initialize workingSet from the feasible initialValues |
| 266 | InequalityFactorGraph workingSet = identifyActiveConstraints( |
| 267 | inequalities: problem_.inequalities, initialValues, duals, useWarmStart); |
| 268 | State state(initialValues, duals, workingSet, false, 0); |
| 269 | |
| 270 | /// main loop of the solver |
| 271 | while (!state.converged) state = iterate(state); |
| 272 | |
| 273 | return std::make_pair(state.values, state.duals); |
| 274 | } |
| 275 | |
| 276 | //****************************************************************************** |
| 277 | Template std::pair<VectorValues, VectorValues> This::optimize() const { |
| 278 | INITSOLVER initSolver(problem_); |
| 279 | VectorValues initValues = initSolver.solve(); |
| 280 | return optimize(initValues); |
| 281 | } |
| 282 | |
| 283 | } |
| 284 | |
| 285 | #undef Template |
| 286 | #undef This |
| 287 | |