potential_optimizer.h

#ifndef POTENTIALS_POTENTIAL_OPTIMIZER_H
#define POTENTIALS_POTENTIAL_OPTIMIZER_H
 
#include "downward/task_proxy.h"
 
#include "downward/lp/lp_solver.h"
 
#include <memory>
#include <vector>
 
namespace potentials {
class PotentialFunction;
 
/*
  Add admissible potential constraints to LP and allow optimizing for
  different objectives.
 
  The implementation is based on transforming the task into transition
  normal form (TNF) (Pommerening and Helmert, ICAPS 2015). We add an
  undefined value u to each variable's domain, but do not create
  forgetting operators or change existing operators explicitly.
  Instead, we ensure that all fact potentials of a variable V are less
  than or equal to the potential of V=u and implicitly set pre(op) to
  V=u for operators op where pre(op) is undefined. Similarly, we set
  s_\star(V) = u for variables V without a goal value.
 
  For more information we refer to the paper introducing potential
  heuristics
 
  Florian Pommerening, Malte Helmert, Gabriele Roeger and Jendrik Seipp.
  From Non-Negative to General Operator Cost Partitioning.
  AAAI 2015.
 
  and the paper comparing various optimization functions
 
  Jendrik Seipp, Florian Pommerening and Malte Helmert.
  New Optimization Functions for Potential Heuristics.
  ICAPS 2015.
*/
class PotentialOptimizer {
    std::shared_ptr<AbstractTask> task;
    TaskProxy task_proxy;
    lp::LPSolver lp_solver;
    const double max_potential;
    int num_lp_vars;
    std::vector<std::vector<int>> lp_var_ids;
    std::vector<std::vector<double>> fact_potentials;
 
    int get_lp_var_id(const FactProxy& fact) const;
    void initialize();
    void construct_lp();
    void solve_and_extract();
    void extract_lp_solution();
 
public:
    PotentialOptimizer(
        const std::shared_ptr<AbstractTask>& transform,
        lp::LPSolverType lpsolver,
        double max_potential);
    ~PotentialOptimizer() = default;
 
    std::shared_ptr<AbstractTask> get_task() const;
    bool potentials_are_bounded() const;
 
    void optimize_for_state(const State& state);
    void optimize_for_all_states();
    void optimize_for_samples(const std::vector<State>& samples);
 
    bool has_optimal_solution() const;
 
    std::unique_ptr<PotentialFunction> get_potential_function() const;
};
} // namespace potentials
 
#endif