Commit a92d3b0c authored by Julien David's avatar Julien David
Browse files

First state revisited

parent ffe85080

src/RandomDistributionGenerator.hpp

+49 −33
Changes for src/RandomDistributionGenerator.hpp: 49 added lines, 33 removed lines.
Original line number Diff line number Diff line
@@ -4,6 +4,8 @@ 
#include"ConcaveFunction.hpp"

#include<assert.h>

#include<unistd.h>

#include<stdexcept>

#include <iomanip>



template<typename F = ConcaveFunction>

class RandomDistributionGenerator{
@@ -19,21 +21,6 @@ private: 


  F cf;



  /**

   Initializes all the metrics of the generator's benchmarks.

   @param nb_xp is the number of conducted experiments

   @ensures all metrics are set to zero, except for the number of experiments

  */

  void init_benchmarks(){

    __AIM_STEPS__ = 0;

    __FIRST_STEP_FAILURES__ = 0;

    __INC_DIST_FAILURES__ = 0;

    __MARKOV_STEP_FAILURES__ = 0;

    __FIRST_STATE_ITERATIONS__ = 0;

    __NB_XP__ = 0;

    __NEWTON_STEPS__ = 0;

    __NEWTON_CALLS__ = 0;

  }



  void distribution_with_max_concave(long double * res, int n){

    int i;
@@ -78,9 +65,10 @@ private: 
    // If the target is unreachable, exit function, only happens when precision error occurs

    //assert(sum<=1);

    if(aim_p < target || cf.aim_double(EPSILON, sum-EPSILON) > target) {

      //std::cout<<"FAILURE "<<aim_p<<" "<<target<<" "<<cf.aim_double(EPSILON, sum-EPSILON)<<std::endl;

      return -1.;

    }

    while (step > EPSILON && (double) aim_p != (double) target && p > 0){

    while ((double)step > (double)EPSILON && (double) aim_p != (double) target && p > 0){

      if(aim_p < target)

	p += step;

      else
@@ -102,52 +90,64 @@ private: 
  void distribution_markov_chain(long double * res, long double target, int size, int steps){

    int i;



    while( reach_first_state(res, target, size) < 0)

    while( reach_first_state(res, target, size) < 0){

      __FIRST_STEP_FAILURES__++;

      if(__FIRST_STEP_FAILURES__ == 10)

	throw (target);

    }



    for (i = 0; i < steps; i++)

      markov_chain_step(res, size);        

  }



  

  

  int reach_first_state(long double * distribution, long double target, int size){

    long double incomplete_sum, contribution;

    long double min_aim, max_aim;

    long double min_aim, max_aim, update;

    long double delta = 0.5/(long double)size;

    distribution_with_max_concave(distribution, size);

    incomplete_sum = distribution[0]*(size-2);

    contribution = cf.contribution(distribution, size-2);

    min_aim = cf.minimal_of_two_values(1-incomplete_sum);

    max_aim = cf.maximal_of_two_values(1-incomplete_sum);

    

    long double update = cf.update_target(target, contribution, size, size-2);



    while( max_aim < update || min_aim > update ){

    distribution[size-2] = (1-incomplete_sum)/2;

    distribution[size-1] = (1-incomplete_sum)/2;

    max_aim = cf.h(cf.contribution(distribution, size), size);

    distribution[size-2] = (1-incomplete_sum-EPSILON);

    distribution[size-1] = EPSILON;

    min_aim = cf.h(cf.contribution(distribution, size), size);

    

    while( (max_aim < target || min_aim > target) && delta > EPSILON ){

      __FIRST_STATE_ITERATIONS__++;



      if( min_aim > target )

	add_delta(distribution, size-2, cf.sign()*delta);

	add_delta(distribution, size-2, -delta);

      else

	add_delta(distribution, size-2, -cf.sign()*delta);

	add_delta(distribution, size-2, delta);

      

      incomplete_sum = distribution[0]*(size-2);

      contribution = cf.contribution(distribution, size-2);

      min_aim = cf.minimal_of_two_values(1-incomplete_sum);

      max_aim = cf.maximal_of_two_values(1-incomplete_sum);

      distribution[size-2] = (1-incomplete_sum)/2;

      distribution[size-1] = (1-incomplete_sum)/2;

      max_aim = cf.h(cf.contribution(distribution, size),size);

      distribution[size-2] = (1-incomplete_sum-EPSILON);

      distribution[size-1] = EPSILON;

      min_aim = cf.h(cf.contribution(distribution, size),size);



      delta/=2;

      update = cf.update_target(target, contribution, size, size-2);

      //std::cout<<distribution[0]<<" "<<min_aim<<" "<<max_aim<<" IT:"<<__FIRST_STATE_ITERATIONS__<<" "<<delta<<std::endl;

      //std::cout<<"x:"<<distribution[0]<<" min:"<<min_aim<<" max:"<<max_aim<<" IT:"<<__FIRST_STATE_ITERATIONS__<<" delta:"<<delta<<std::endl;

    }

    contribution = cf.contribution(distribution, size-2);

    update = cf.update_target(target, contribution, size, size-2);

    distribution[size-2] = aim_concave_value(1-incomplete_sum, update);

    

    if(distribution[size-2] < 0)

      return -1;

  										

    distribution[size-1] = 1-incomplete_sum - distribution[size-2];

    //print_distribution(distribution, size);

    return 0; //If equal to -1, then a problem occured.

  }





  



  void markov_chain_step(long double * dist, int k){

    int triplet[3], i;

    long double target;
@@ -206,6 +206,22 @@ public: 
      std::cerr<<"The target is too big to be reached"<<std::endl;

  }



  /**

   Initializes all the metrics of the generator's benchmarks.

   @param nb_xp is the number of conducted experiments

   @ensures all metrics are set to zero, except for the number of experiments

  */

  void init_benchmarks(){

    __AIM_STEPS__ = 0;

    __FIRST_STEP_FAILURES__ = 0;

    __INC_DIST_FAILURES__ = 0;

    __MARKOV_STEP_FAILURES__ = 0;

    __FIRST_STATE_ITERATIONS__ = 0;

    __NB_XP__ = 0;

    __NEWTON_STEPS__ = 0;

    __NEWTON_CALLS__ = 0;

  }





  void print_benchmarks(){

    print_benchmarks(__NB_XP__);