alm-helpers.hpp

Go to the documentation of this file.

#pragma once
 
#include <alpaqa/decl/alm.hpp>
#include <stdexcept>
 
namespace alpaqa::detail {
 
inline void project_y(rvec y,     // inout
                      crvec z_lb, // in
                      crvec z_ub, // in
                      real_t M    // in
) {
    // TODO: Handle NaN correctly
    auto max_lb = [M](real_t y, real_t z_lb) {
        real_t y_lb = z_lb == -inf ? 0 : -M;
        return std::max(y, y_lb);
    };
    auto min_ub = [M](real_t y, real_t z_ub) {
        real_t y_ub = z_ub == inf ? 0 : M;
        return std::min(y, y_ub);
    };
    y = y.binaryExpr(z_lb, max_lb).binaryExpr(z_ub, min_ub);
}
 
inline void update_penalty_weights(const ALMParams &params, real_t Δ,
                                   bool first_iter, rvec e, rvec old_e,
                                   real_t norm_e, real_t old_norm_e,
                                   crvec Σ_old, rvec Σ) {
    if (norm_e <= params.δ) {
        Σ = Σ_old;
        return;
    }
    if (params.single_penalty_factor) {
        if (first_iter || norm_e > params.θ * old_norm_e) {
            real_t new_Σ = std::fmin(params.Σ_max, Δ * Σ_old(0));
            Σ.setConstant(new_Σ);
        } else {
            Σ = Σ_old;
        }
    } else {
        for (Eigen::Index i = 0; i < e.rows(); ++i) {
            if (first_iter || std::abs(e(i)) > params.θ * std::abs(old_e(i))) {
                Σ(i) = std::fmin(params.Σ_max,
                                 std::fmax(Δ * std::abs(e(i)) / norm_e, 1) *
                                     Σ_old(i));
            } else {
                Σ(i) = Σ_old(i);
            }
        }
    }
}
 
inline void initialize_penalty(const Problem &p, const ALMParams &params,
                               crvec x0, rvec Σ) {
    real_t f0 = p.f(x0);
    vec g0(p.m);
    p.g(x0, g0);
    // TODO: reuse evaluations of f ang g in PANOC?
    real_t σ = params.σ₀ * std::max(real_t(1), std::abs(f0)) /
               std::max(real_t(1), 0.5 * g0.squaredNorm());
    σ = std::max(σ, params.Σ_min);
    σ = std::min(σ, params.Σ_max);
    Σ.fill(σ);
}
 
inline void apply_preconditioning(const Problem &problem, Problem &prec_problem,
                                  crvec x, real_t &prec_f, vec &prec_g) {
    vec grad_f(problem.n);
    vec v = vec::Zero(problem.m);
    vec grad_g(problem.n);
    problem.grad_f(x, grad_f);
    prec_f = 1. / std::max(grad_f.lpNorm<Eigen::Infinity>(), real_t{1});
    prec_g.resize(problem.m);
 
    for (Eigen::Index i = 0; i < problem.m; ++i) {
        v(i) = 1;
        problem.grad_g_prod(x, v, grad_g);
        v(i)      = 0;
        prec_g(i) = 1. / std::max(grad_g.lpNorm<Eigen::Infinity>(), real_t{1});
    }
 
    auto prec_f_fun      = [&](crvec x) { return problem.f(x) * prec_f; };
    auto prec_grad_f_fun = [&](crvec x, rvec grad_f) {
        problem.grad_f(x, grad_f);
        grad_f *= prec_f;
    };
    auto prec_g_fun = [&](crvec x, rvec g) {
        problem.g(x, g);
        g = prec_g.asDiagonal() * g;
    };
    auto prec_grad_g_fun = [&](crvec x, crvec v, rvec grad_g) {
        vec prec_v = prec_g.asDiagonal() * v;
        problem.grad_g_prod(x, prec_v, grad_g);
    };
    prec_problem = Problem{
        problem.n,
        problem.m,
        problem.C,
        Box{prec_g.asDiagonal() * problem.D.upperbound,
            prec_g.asDiagonal() * problem.D.lowerbound},
        std::move(prec_f_fun),
        std::move(prec_grad_f_fun),
        std::move(prec_g_fun),
        std::move(prec_grad_g_fun),
        [](crvec, unsigned, rvec) {
            throw std::logic_error("Preconditioning for second-order solvers "
                                   "has not yet been implemented");
        },
        [](crvec, crvec, crvec, rvec) {
            throw std::logic_error("Preconditioning for second-order solvers "
                                   "has not yet been implemented");
        },
        [](crvec, crvec, rmat) {
            throw std::logic_error("Preconditioning for second-order solvers "
                                   "has not yet been implemented");
        },
    };
}
 
} // namespace alpaqa::detail