Austin Schuh | 2944103 | 2023-05-31 19:32:24 -0700 | [diff] [blame] | 1 | #ifndef FRC971_SOLVERS_CONVEX_H_ |
| 2 | #define FRC971_SOLVERS_CONVEX_H_ |
| 3 | |
Austin Schuh | e9ae456 | 2023-04-25 22:18:18 -0700 | [diff] [blame] | 4 | #include <sys/types.h> |
| 5 | #include <unistd.h> |
| 6 | |
Austin Schuh | e9ae456 | 2023-04-25 22:18:18 -0700 | [diff] [blame] | 7 | #include <iomanip> |
| 8 | |
Austin Schuh | 99f7c6a | 2024-06-25 22:07:44 -0700 | [diff] [blame] | 9 | #include "absl/log/check.h" |
| 10 | #include "absl/log/log.h" |
Austin Schuh | e9ae456 | 2023-04-25 22:18:18 -0700 | [diff] [blame] | 11 | #include "absl/strings/str_join.h" |
Philipp Schrader | 790cb54 | 2023-07-05 21:06:52 -0700 | [diff] [blame] | 12 | #include <Eigen/Dense> |
Austin Schuh | e9ae456 | 2023-04-25 22:18:18 -0700 | [diff] [blame] | 13 | |
Stephan Pleines | d99b1ee | 2024-02-02 20:56:44 -0800 | [diff] [blame] | 14 | namespace frc971::solvers { |
Austin Schuh | e9ae456 | 2023-04-25 22:18:18 -0700 | [diff] [blame] | 15 | |
| 16 | // TODO(austin): Steal JET from Ceres to generate the derivatives easily and |
| 17 | // quickly? |
| 18 | // |
| 19 | // States is the number of inputs to the optimization problem. |
| 20 | // M is the number of inequality constraints. |
| 21 | // N is the number of equality constraints. |
| 22 | template <size_t States, size_t M, size_t N> |
| 23 | class ConvexProblem { |
| 24 | public: |
| 25 | // Returns the function to minimize and it's derivatives. |
| 26 | virtual double f0( |
| 27 | Eigen::Ref<const Eigen::Matrix<double, States, 1>> X) const = 0; |
| 28 | virtual Eigen::Matrix<double, States, 1> df0( |
| 29 | Eigen::Ref<const Eigen::Matrix<double, States, 1>> X) const = 0; |
| 30 | virtual Eigen::Matrix<double, States, States> ddf0( |
| 31 | Eigen::Ref<const Eigen::Matrix<double, States, 1>> X) const = 0; |
| 32 | |
| 33 | // Returns the constraints f(X) < 0, and their derivative. |
| 34 | virtual Eigen::Matrix<double, M, 1> f( |
| 35 | Eigen::Ref<const Eigen::Matrix<double, States, 1>> X) const = 0; |
| 36 | virtual Eigen::Matrix<double, M, States> df( |
| 37 | Eigen::Ref<const Eigen::Matrix<double, States, 1>> X) const = 0; |
| 38 | |
| 39 | // Returns the equality constraints of the form A x = b |
| 40 | virtual Eigen::Matrix<double, N, States> A() const = 0; |
| 41 | virtual Eigen::Matrix<double, N, 1> b() const = 0; |
| 42 | }; |
| 43 | |
| 44 | // Implements a Primal-Dual Interior point method convex solver. |
| 45 | // See 11.7 of https://web.stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf |
| 46 | // |
| 47 | // States is the number of inputs to the optimization problem. |
| 48 | // M is the number of inequality constraints. |
| 49 | // N is the number of equality constraints. |
| 50 | template <size_t States, size_t M, size_t N> |
| 51 | class Solver { |
| 52 | public: |
| 53 | // Ratio to require the cost to decrease when line searching. |
| 54 | static constexpr double kAlpha = 0.05; |
| 55 | // Line search step parameter. |
| 56 | static constexpr double kBeta = 0.5; |
| 57 | static constexpr double kMu = 2.0; |
| 58 | // Terminal condition for the primal problem (equality constraints) and dual |
| 59 | // (gradient + inequality constraints). |
| 60 | static constexpr double kEpsilonF = 1e-6; |
| 61 | // Terminal condition for nu, the surrogate duality gap. |
| 62 | static constexpr double kEpsilon = 1e-6; |
| 63 | |
| 64 | // Solves the problem given a feasible initial solution. |
| 65 | Eigen::Matrix<double, States, 1> Solve( |
| 66 | const ConvexProblem<States, M, N> &problem, |
| 67 | Eigen::Ref<const Eigen::Matrix<double, States, 1>> X_initial); |
| 68 | |
| 69 | private: |
| 70 | // Class to hold all the derivataves and function evaluations. |
| 71 | struct Derivatives { |
| 72 | Eigen::Matrix<double, States, 1> gradient; |
| 73 | Eigen::Matrix<double, States, States> hessian; |
| 74 | |
| 75 | // Inequality function f |
| 76 | Eigen::Matrix<double, M, 1> f; |
| 77 | // df |
| 78 | Eigen::Matrix<double, M, States> df; |
| 79 | |
| 80 | // ddf is assumed to be 0 because for the linear constraint distance |
| 81 | // function we are using, it is actually 0, and by assuming it is zero |
| 82 | // rather than passing it through as 0 to the solver, we can save enough CPU |
| 83 | // to make it worth it. |
| 84 | |
| 85 | // A |
| 86 | Eigen::Matrix<double, N, States> A; |
| 87 | // Ax - b |
| 88 | Eigen::Matrix<double, N, 1> Axmb; |
| 89 | }; |
| 90 | |
| 91 | // Computes all the values for the given problem at the given state. |
| 92 | Derivatives ComputeDerivative( |
| 93 | const ConvexProblem<States, M, N> &problem, |
| 94 | const Eigen::Ref<const Eigen::Matrix<double, States + M + N, 1>> y); |
| 95 | |
| 96 | // Computes Rt at the given state and with the given t_inverse. See 11.53 of |
| 97 | // cvxbook.pdf. |
| 98 | Eigen::Matrix<double, States + M + N, 1> Rt( |
| 99 | const Derivatives &derivatives, |
| 100 | Eigen::Matrix<double, States + M + N, 1> y, double t_inverse); |
| 101 | |
| 102 | // Prints out all the derivatives with VLOG at the provided verbosity. |
| 103 | void PrintDerivatives( |
| 104 | const Derivatives &derivatives, |
| 105 | const Eigen::Ref<const Eigen::Matrix<double, States + M + N, 1>> y, |
| 106 | std::string_view prefix, int verbosity); |
| 107 | }; |
| 108 | |
| 109 | template <size_t States, size_t M, size_t N> |
| 110 | Eigen::Matrix<double, States + M + N, 1> Solver<States, M, N>::Rt( |
| 111 | const Derivatives &derivatives, Eigen::Matrix<double, States + M + N, 1> y, |
| 112 | double t_inverse) { |
| 113 | Eigen::Matrix<double, States + M + N, 1> result; |
| 114 | |
| 115 | Eigen::Ref<Eigen::Matrix<double, States, 1>> r_dual = |
| 116 | result.template block<States, 1>(0, 0); |
| 117 | Eigen::Ref<Eigen::Matrix<double, M, 1>> r_cent = |
| 118 | result.template block<M, 1>(States, 0); |
| 119 | Eigen::Ref<Eigen::Matrix<double, N, 1>> r_pri = |
| 120 | result.template block<N, 1>(States + M, 0); |
| 121 | |
| 122 | Eigen::Ref<const Eigen::Matrix<double, M, 1>> lambda = |
| 123 | y.template block<M, 1>(States, 0); |
| 124 | Eigen::Ref<const Eigen::Matrix<double, N, 1>> v = |
| 125 | y.template block<N, 1>(States + M, 0); |
| 126 | |
| 127 | r_dual = derivatives.gradient + derivatives.df.transpose() * lambda + |
| 128 | derivatives.A.transpose() * v; |
Austin Schuh | 2944103 | 2023-05-31 19:32:24 -0700 | [diff] [blame] | 129 | r_cent = -lambda.array() * derivatives.f.array() - t_inverse; |
Austin Schuh | e9ae456 | 2023-04-25 22:18:18 -0700 | [diff] [blame] | 130 | r_pri = derivatives.Axmb; |
| 131 | |
| 132 | return result; |
| 133 | } |
| 134 | |
| 135 | template <size_t States, size_t M, size_t N> |
| 136 | Eigen::Matrix<double, States, 1> Solver<States, M, N>::Solve( |
| 137 | const ConvexProblem<States, M, N> &problem, |
| 138 | Eigen::Ref<const Eigen::Matrix<double, States, 1>> X_initial) { |
| 139 | const Eigen::IOFormat kHeavyFormat(Eigen::StreamPrecision, 0, ", ", |
| 140 | ",\n " |
| 141 | " ", |
| 142 | "[", "]", "[", "]"); |
| 143 | |
| 144 | Eigen::Matrix<double, States + M + N, 1> y = |
| 145 | Eigen::Matrix<double, States + M + N, 1>::Constant(1.0); |
| 146 | y.template block<States, 1>(0, 0) = X_initial; |
| 147 | |
| 148 | Derivatives derivatives = ComputeDerivative(problem, y); |
| 149 | |
| 150 | for (size_t i = 0; i < M; ++i) { |
| 151 | CHECK_LE(derivatives.f(i, 0), 0.0) |
| 152 | << ": Initial state " << X_initial.transpose().format(kHeavyFormat) |
| 153 | << " not feasible"; |
| 154 | } |
| 155 | |
| 156 | PrintDerivatives(derivatives, y, "", 1); |
| 157 | |
| 158 | size_t iteration = 0; |
| 159 | while (true) { |
| 160 | // Solve for the primal-dual search direction by solving the newton step. |
| 161 | Eigen::Ref<const Eigen::Matrix<double, M, 1>> lambda = |
| 162 | y.template block<M, 1>(States, 0); |
| 163 | |
| 164 | const double nu = -(derivatives.f.transpose() * lambda)(0, 0); |
| 165 | const double t_inverse = nu / (kMu * lambda.rows()); |
| 166 | Eigen::Matrix<double, States + M + N, 1> rt_orig = |
| 167 | Rt(derivatives, y, t_inverse); |
| 168 | |
| 169 | Eigen::Matrix<double, States + M + N, States + M + N> m1; |
| 170 | m1.setZero(); |
| 171 | m1.template block<States, States>(0, 0) = derivatives.hessian; |
| 172 | m1.template block<States, M>(0, States) = derivatives.df.transpose(); |
| 173 | m1.template block<States, N>(0, States + M) = derivatives.A.transpose(); |
| 174 | m1.template block<M, States>(States, 0) = |
| 175 | -(Eigen::DiagonalMatrix<double, M>(lambda) * derivatives.df); |
Austin Schuh | 2944103 | 2023-05-31 19:32:24 -0700 | [diff] [blame] | 176 | m1.template block<M, M>(States, States) = |
| 177 | Eigen::DiagonalMatrix<double, M>(-derivatives.f); |
Austin Schuh | e9ae456 | 2023-04-25 22:18:18 -0700 | [diff] [blame] | 178 | m1.template block<N, States>(States + M, 0) = derivatives.A; |
| 179 | |
| 180 | Eigen::Matrix<double, States + M + N, 1> dy = |
| 181 | m1.colPivHouseholderQr().solve(-rt_orig); |
| 182 | |
| 183 | Eigen::Ref<Eigen::Matrix<double, M, 1>> dlambda = |
| 184 | dy.template block<M, 1>(States, 0); |
| 185 | |
| 186 | double s = 1.0; |
| 187 | |
| 188 | // Now, time to do line search. |
| 189 | // |
| 190 | // Start by keeping lambda positive. Make sure our step doesn't let |
| 191 | // lambda cross 0. |
| 192 | for (int i = 0; i < dlambda.rows(); ++i) { |
| 193 | if (lambda(i) + s * dlambda(i) < 0.0) { |
| 194 | // Ignore tiny steps in lambda. They cause issues when we get really |
| 195 | // close to having our constraints met but haven't converged the rest |
| 196 | // of the problem and start to run into rounding issues in the matrix |
| 197 | // solve portion. |
| 198 | if (dlambda(i) < 0.0 && dlambda(i) > -1e-12) { |
| 199 | VLOG(1) << " lambda(" << i << ") " << lambda(i) << " + " << s |
| 200 | << " * " << dlambda(i) << " -> s would be now " |
| 201 | << -lambda(i) / dlambda(i); |
| 202 | dlambda(i) = 0.0; |
| 203 | VLOG(1) << " dy -> " << std::setprecision(12) << std::fixed |
| 204 | << std::setfill(' ') << dy.transpose().format(kHeavyFormat); |
| 205 | continue; |
| 206 | } |
| 207 | VLOG(1) << " lambda(" << i << ") " << lambda(i) << " + " << s << " * " |
| 208 | << dlambda(i) << " -> s now " << -lambda(i) / dlambda(i); |
| 209 | s = -lambda(i) / dlambda(i); |
| 210 | } |
| 211 | } |
| 212 | |
| 213 | VLOG(1) << " After lambda line search, s is " << s; |
| 214 | |
| 215 | VLOG(3) << " Initial step " << iteration << " -> " << std::setprecision(12) |
| 216 | << std::fixed << std::setfill(' ') |
| 217 | << dy.transpose().format(kHeavyFormat); |
| 218 | VLOG(3) << " rt -> " |
| 219 | << std::setprecision(12) << std::fixed << std::setfill(' ') |
| 220 | << rt_orig.transpose().format(kHeavyFormat); |
| 221 | |
| 222 | const double rt_orig_squared_norm = rt_orig.squaredNorm(); |
| 223 | |
| 224 | Eigen::Matrix<double, States + M + N, 1> next_y; |
| 225 | Eigen::Matrix<double, States + M + N, 1> rt; |
| 226 | Derivatives next_derivatives; |
| 227 | while (true) { |
| 228 | next_y = y + s * dy; |
| 229 | next_derivatives = ComputeDerivative(problem, next_y); |
| 230 | rt = Rt(next_derivatives, next_y, t_inverse); |
| 231 | |
| 232 | const Eigen::Ref<const Eigen::VectorXd> next_x = |
| 233 | next_y.block(0, 0, next_derivatives.hessian.rows(), 1); |
| 234 | const Eigen::Ref<const Eigen::VectorXd> next_lambda = |
| 235 | next_y.block(next_x.rows(), 0, next_derivatives.f.rows(), 1); |
| 236 | |
| 237 | const Eigen::Ref<const Eigen::VectorXd> next_v = next_y.block( |
| 238 | next_x.rows() + next_lambda.rows(), 0, next_derivatives.A.rows(), 1); |
| 239 | |
| 240 | VLOG(1) << " next_rt(" << iteration << ") is " << rt.norm() << " -> " |
| 241 | << std::setprecision(12) << std::fixed << std::setfill(' ') |
| 242 | << rt.transpose().format(kHeavyFormat); |
| 243 | |
| 244 | PrintDerivatives(next_derivatives, next_y, "next_", 3); |
| 245 | |
| 246 | if (next_derivatives.f.maxCoeff() > 0.0) { |
| 247 | VLOG(1) << " f_next > 0.0 -> " << next_derivatives.f.maxCoeff() |
| 248 | << ", continuing line search."; |
| 249 | s *= kBeta; |
| 250 | } else if (next_derivatives.Axmb.squaredNorm() < 0.1 && |
| 251 | rt.squaredNorm() > |
| 252 | std::pow(1.0 - kAlpha * s, 2.0) * rt_orig_squared_norm) { |
| 253 | VLOG(1) << " |Rt| > |Rt+1| " << rt.norm() << " > " << rt_orig.norm() |
| 254 | << ", drt -> " << std::setprecision(12) << std::fixed |
| 255 | << std::setfill(' ') |
| 256 | << (rt_orig - rt).transpose().format(kHeavyFormat); |
| 257 | s *= kBeta; |
| 258 | } else { |
| 259 | break; |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | VLOG(1) << " Terminated line search with s " << s << ", " << rt.norm() |
| 264 | << "(|Rt+1|) < " << rt_orig.norm() << "(|Rt|)"; |
| 265 | y = next_y; |
| 266 | |
| 267 | const Eigen::Ref<const Eigen::VectorXd> next_lambda = |
| 268 | y.template block<M, 1>(States, 0); |
| 269 | |
| 270 | // See if we hit our convergence criteria. |
| 271 | const double r_primal_squared_norm = |
| 272 | rt.template block<N, 1>(States + M, 0).squaredNorm(); |
| 273 | VLOG(1) << " rt_next(" << iteration << ") is " << rt.norm() << " -> " |
| 274 | << std::setprecision(12) << std::fixed << std::setfill(' ') |
| 275 | << rt.transpose().format(kHeavyFormat); |
| 276 | if (r_primal_squared_norm < kEpsilonF * kEpsilonF) { |
| 277 | const double r_dual_squared_norm = |
| 278 | rt.template block<States, 1>(0, 0).squaredNorm(); |
| 279 | if (r_dual_squared_norm < kEpsilonF * kEpsilonF) { |
| 280 | const double next_nu = |
| 281 | -(next_derivatives.f.transpose() * next_lambda)(0, 0); |
| 282 | if (next_nu < kEpsilon) { |
| 283 | VLOG(1) << " r_primal(" << iteration << ") -> " |
| 284 | << std::sqrt(r_primal_squared_norm) << " < " << kEpsilonF |
| 285 | << ", r_dual(" << iteration << ") -> " |
| 286 | << std::sqrt(r_dual_squared_norm) << " < " << kEpsilonF |
| 287 | << ", nu(" << iteration << ") -> " << next_nu << " < " |
| 288 | << kEpsilon; |
| 289 | break; |
| 290 | } else { |
| 291 | VLOG(1) << " nu(" << iteration << ") -> " << next_nu << " < " |
| 292 | << kEpsilon << ", not done yet"; |
| 293 | } |
| 294 | |
| 295 | } else { |
| 296 | VLOG(1) << " r_dual(" << iteration << ") -> " |
| 297 | << std::sqrt(r_dual_squared_norm) << " < " << kEpsilonF |
| 298 | << ", not done yet"; |
| 299 | } |
| 300 | } else { |
| 301 | VLOG(1) << " r_primal(" << iteration << ") -> " |
| 302 | << std::sqrt(r_primal_squared_norm) << " < " << kEpsilonF |
| 303 | << ", not done yet"; |
| 304 | } |
| 305 | VLOG(1) << " step(" << iteration << ") " << std::setprecision(12) |
| 306 | << (s * dy).transpose().format(kHeavyFormat); |
| 307 | VLOG(1) << " y(" << iteration << ") is now " << std::setprecision(12) |
| 308 | << y.transpose().format(kHeavyFormat); |
| 309 | |
| 310 | // Very import, use the last set of derivatives we picked for our new y |
| 311 | // for the next iteration. This avoids re-computing it. |
| 312 | derivatives = std::move(next_derivatives); |
| 313 | |
| 314 | ++iteration; |
| 315 | if (iteration > 100) { |
| 316 | LOG(FATAL) << "Too many iterations"; |
| 317 | } |
| 318 | } |
| 319 | |
| 320 | return y.template block<States, 1>(0, 0); |
| 321 | } |
| 322 | |
| 323 | template <size_t States, size_t M, size_t N> |
| 324 | typename Solver<States, M, N>::Derivatives |
| 325 | Solver<States, M, N>::ComputeDerivative( |
| 326 | const ConvexProblem<States, M, N> &problem, |
| 327 | const Eigen::Ref<const Eigen::Matrix<double, States + M + N, 1>> y) { |
| 328 | const Eigen::Ref<const Eigen::Matrix<double, States, 1>> x = |
| 329 | y.template block<States, 1>(0, 0); |
| 330 | |
| 331 | Derivatives derivatives; |
| 332 | derivatives.gradient = problem.df0(x); |
| 333 | derivatives.hessian = problem.ddf0(x); |
| 334 | derivatives.f = problem.f(x); |
| 335 | derivatives.df = problem.df(x); |
| 336 | derivatives.A = problem.A(); |
| 337 | derivatives.Axmb = |
| 338 | derivatives.A * y.template block<States, 1>(0, 0) - problem.b(); |
| 339 | return derivatives; |
| 340 | } |
| 341 | |
| 342 | template <size_t States, size_t M, size_t N> |
| 343 | void Solver<States, M, N>::PrintDerivatives( |
| 344 | const Derivatives &derivatives, |
| 345 | const Eigen::Ref<const Eigen::Matrix<double, States + M + N, 1>> y, |
| 346 | std::string_view prefix, int verbosity) { |
| 347 | const Eigen::Ref<const Eigen::VectorXd> x = |
| 348 | y.block(0, 0, derivatives.hessian.rows(), 1); |
| 349 | const Eigen::Ref<const Eigen::VectorXd> lambda = |
| 350 | y.block(x.rows(), 0, derivatives.f.rows(), 1); |
| 351 | |
| 352 | if (VLOG_IS_ON(verbosity)) { |
| 353 | Eigen::IOFormat heavy(Eigen::StreamPrecision, 0, ", ", |
| 354 | ",\n " |
| 355 | " ", |
| 356 | "[", "]", "[", "]"); |
| 357 | heavy.rowSeparator = |
| 358 | heavy.rowSeparator + |
| 359 | std::string(absl::StrCat(getpid()).size() + prefix.size(), ' '); |
| 360 | |
| 361 | const Eigen::Ref<const Eigen::VectorXd> v = |
| 362 | y.block(x.rows() + lambda.rows(), 0, derivatives.A.rows(), 1); |
| 363 | VLOG(verbosity) << " " << prefix << "x: " << x.transpose().format(heavy); |
| 364 | VLOG(verbosity) << " " << prefix |
| 365 | << "lambda: " << lambda.transpose().format(heavy); |
| 366 | VLOG(verbosity) << " " << prefix << "v: " << v.transpose().format(heavy); |
| 367 | VLOG(verbosity) << " " << prefix |
| 368 | << "hessian: " << derivatives.hessian.format(heavy); |
| 369 | VLOG(verbosity) << " " << prefix |
| 370 | << "gradient: " << derivatives.gradient.format(heavy); |
| 371 | VLOG(verbosity) << " " << prefix |
| 372 | << "A: " << derivatives.A.format(heavy); |
| 373 | VLOG(verbosity) << " " << prefix |
| 374 | << "Ax-b: " << derivatives.Axmb.format(heavy); |
| 375 | VLOG(verbosity) << " " << prefix |
| 376 | << "f: " << derivatives.f.format(heavy); |
| 377 | VLOG(verbosity) << " " << prefix |
| 378 | << "df: " << derivatives.df.format(heavy); |
| 379 | } |
| 380 | } |
| 381 | |
Stephan Pleines | d99b1ee | 2024-02-02 20:56:44 -0800 | [diff] [blame] | 382 | } // namespace frc971::solvers |
Austin Schuh | 2944103 | 2023-05-31 19:32:24 -0700 | [diff] [blame] | 383 | |
| 384 | #endif // FRC971_SOLVERS_CONVEX_H_ |