lemon/simann.h
author deba
Tue, 31 Jan 2006 20:04:36 +0000
changeset 1933 a876a3d6a4c7
parent 1918 09415ae11103
child 1956 a055123339d5
permissions -rw-r--r--
Revising the bpugraph concept

We need a public but very limited ANode and BNode class
It can be used with ItemSetTraits and with some special maps

By example:
DescriptorMap<Graph, ANode>
InvertableMap<Graph, ANode, string>
IterableBoolMap<Graph, ANode>
IterableIntMap<Graph, ANode>
IterableValueMap<Graph, ANode, string>
     1 // -*- C++ -*-
     2 #ifndef LEMON_SIMANN_H
     3 #define LEMON_SIMANN_H
     4 
     5 /// \ingroup experimental
     6 /// \file
     7 /// \brief Simulated annealing framework.
     8 ///
     9 /// \todo A test and some demo should be added
    10 /// \todo Doc should be improved
    11 /// \author Akos Ladanyi
    12 
    13 #include <cstdlib>
    14 #include <cmath>
    15 #include <limits>
    16 #include <lemon/time_measure.h>
    17 
    18 namespace lemon {
    19 
    20 /// \addtogroup experimental
    21 /// @{
    22 
    23   class SimAnnBase;
    24 
    25   /// \brief A base class for controllers.
    26   class ControllerBase {
    27   public:
    28     friend class SimAnnBase;
    29     /// \brief Pointer to the simulated annealing base class.
    30     SimAnnBase *simann;
    31     /// \brief Initializes the controller.
    32     virtual void init() {}
    33     /// \brief This is called by the simulated annealing class when a
    34     /// neighbouring state gets accepted.
    35     virtual void acceptEvent() {}
    36     /// \brief This is called by the simulated annealing class when the
    37     /// accepted neighbouring state's cost is less than the best found one's.
    38     virtual void improveEvent() {}
    39     /// \brief This is called by the simulated annealing class when a
    40     /// neighbouring state gets rejected.
    41     virtual void rejectEvent() {}
    42     /// \brief Decides whether to continue the annealing process or not.
    43     virtual bool next() = 0;
    44     /// \brief Decides whether to accept the current solution or not.
    45     virtual bool accept() = 0;
    46     /// \brief Destructor.
    47     virtual ~ControllerBase() {}
    48   };
    49 
    50   /// \brief Skeleton of an entity class.
    51   class EntityBase {
    52   public:
    53     /// \brief Makes a minor change to the entity.
    54     /// \return the new cost
    55     virtual double mutate() = 0;
    56     /// \brief Restores the entity to its previous state i.e. reverts the
    57     /// effects of the last mutate().
    58     virtual void revert() = 0;
    59     /// \brief Makes a copy of the entity.
    60     virtual EntityBase* clone() = 0;
    61     /// \brief Makes a major change to the entity.
    62     virtual void randomize() = 0;
    63     /// \brief Destructor.
    64     virtual ~EntityBase() {}
    65   };
    66 
    67   /// \brief Simulated annealing abstract base class.
    68   /// Can be used to derive a custom simulated annealing class if \ref SimAnn
    69   /// doesn't fit your needs.
    70   class SimAnnBase {
    71   private:
    72     /// \brief Pointer to the controller.
    73     ControllerBase *controller;
    74     /// \brief Cost of the current solution.
    75     double curr_cost;
    76     /// \brief Cost of the best solution.
    77     double best_cost;
    78     /// \brief Cost of the previous solution.
    79     double prev_cost;
    80     /// \brief Cost of the solution preceding the previous one.
    81     double prev_prev_cost;
    82     /// \brief Number of iterations.
    83     long iter;
    84     /// \brief Number of iterations which did not improve the solution since
    85     /// the last improvement.
    86     long last_impr;
    87   protected:
    88     /// \brief Step to a neighbouring state.
    89     virtual double mutate() = 0;
    90     /// \brief Reverts the last mutate().
    91     virtual void revert() = 0;
    92     /// \brief Saves the current solution as the best one.
    93     virtual void saveAsBest() = 0;
    94     /// \brief Does initializations before each run.
    95     virtual void init() {
    96       controller->init();
    97       curr_cost = prev_cost = prev_prev_cost = best_cost =
    98         std::numeric_limits<double>::infinity();
    99       iter = last_impr = 0;
   100     }
   101   public:
   102     /// \brief Sets the controller class to use.
   103     void setController(ControllerBase &_controller) {
   104       controller = &_controller;
   105       controller->simann = this;
   106     }
   107     /// \brief Returns the cost of the current solution.
   108     double getCurrCost() const { return curr_cost; }
   109     /// \brief Returns the cost of the previous solution.
   110     double getPrevCost() const { return prev_cost; }
   111     /// \brief Returns the cost of the best solution.
   112     double getBestCost() const { return best_cost; }
   113     /// \brief Returns the number of iterations done.
   114     long getIter() const { return iter; }
   115     /// \brief Returns the ordinal number of the last iteration when the
   116     /// solution was improved.
   117     long getLastImpr() const { return last_impr; }
   118     /// \brief Performs one iteration.
   119     bool step() {
   120       iter++;
   121       prev_prev_cost = prev_cost;
   122       prev_cost = curr_cost;
   123       curr_cost = mutate();
   124       if (controller->accept()) {
   125         controller->acceptEvent();
   126         last_impr = iter;
   127         if (curr_cost < best_cost) {
   128           best_cost = curr_cost;
   129           saveAsBest();
   130           controller->improveEvent();
   131         }
   132       }
   133       else {
   134         revert();
   135         curr_cost = prev_cost;
   136         prev_cost = prev_prev_cost;
   137         controller->rejectEvent();
   138       }
   139       return controller->next();
   140     }
   141     /// \brief Performs a given number of iterations.
   142     /// \param n the number of iterations
   143     bool step(int n) {
   144       for(; n > 0 && step(); --n) ;
   145       return !n;
   146     }
   147     /// \brief Starts the annealing process.
   148     void run() {
   149       init();
   150       do { } while (step());
   151     }
   152     /// \brief Destructor.
   153     virtual ~SimAnnBase() {}
   154   };
   155 
   156   /// \brief Simulated annealing class.
   157   class SimAnn : public SimAnnBase {
   158   private:
   159     /// \brief Pointer to the current entity.
   160     EntityBase *curr_ent;
   161     /// \brief Pointer to the best entity.
   162     EntityBase *best_ent;
   163     /// \brief Does initializations before each run.
   164     void init() {
   165       SimAnnBase::init();
   166       if (best_ent) delete best_ent;
   167       best_ent = NULL;
   168       curr_ent->randomize();
   169     }
   170   public:
   171     /// \brief Constructor.
   172     SimAnn() : curr_ent(NULL), best_ent(NULL) {}
   173     /// \brief Destructor.
   174     virtual ~SimAnn() {
   175       if (best_ent) delete best_ent;
   176     }
   177     /// \brief Step to a neighbouring state.
   178     double mutate() {
   179       return curr_ent->mutate();
   180     }
   181     /// \brief Reverts the last mutate().
   182     void revert() {
   183       curr_ent->revert();
   184     }
   185     /// \brief Saves the current solution as the best one.
   186     void saveAsBest() { 
   187       if (best_ent) delete best_ent;
   188       best_ent = curr_ent->clone();
   189     }
   190     /// \brief Sets the current entity.
   191     void setEntity(EntityBase &_ent) {
   192       curr_ent = &_ent;
   193     }
   194     /// \brief Returns a copy of the best found entity.
   195     EntityBase* getBestEntity() { return best_ent->clone(); }
   196   };
   197 
   198   /// \brief A simple controller for the simulated annealing class.
   199   /// This controller starts from a given initial temperature and evenly
   200   /// decreases it.
   201   class SimpleController : public ControllerBase {
   202   private:
   203     /// \brief Maximum number of iterations.
   204     long max_iter;
   205     /// \brief Maximum number of iterations which do not improve the
   206     /// solution.
   207     long max_no_impr;
   208     /// \brief Temperature.
   209     double temp;
   210     /// \brief Annealing factor.
   211     double ann_fact;
   212     /// \brief Constructor.
   213     /// \param _max_iter maximum number of iterations
   214     /// \param _max_no_impr maximum number of consecutive iterations which do
   215     ///        not yield a better solution
   216     /// \param _temp initial temperature
   217     /// \param _ann_fact annealing factor
   218   public:
   219     SimpleController(long _max_iter = 500000, long _max_no_impr = 20000,
   220     double _temp = 1000.0, double _ann_fact = 0.9999) : max_iter(_max_iter),
   221       max_no_impr(_max_no_impr), temp(_temp), ann_fact(_ann_fact)
   222     {
   223       srand48(time(0));
   224     }
   225     /// \brief This is called when a neighbouring state gets accepted.
   226     void acceptEvent() {}
   227     /// \brief This is called when the accepted neighbouring state's cost is
   228     /// less than the best found one's.
   229     void improveEvent() {}
   230     /// \brief This is called when a neighbouring state gets rejected.
   231     void rejectEvent() {}
   232     /// \brief Decides whether to continue the annealing process or not. Also
   233     /// decreases the temperature.
   234     bool next() {
   235       temp *= ann_fact;
   236       bool quit = (simann->getIter() > max_iter) ||
   237         (simann->getIter() - simann->getLastImpr() > max_no_impr);
   238       return !quit;
   239     }
   240     /// \brief Decides whether to accept the current solution or not.
   241     bool accept() {
   242       double cost_diff = simann->getCurrCost() - simann->getPrevCost();
   243       return (drand48() <= exp(-(cost_diff / temp)));
   244     }
   245     /// \brief Destructor.
   246     virtual ~SimpleController() {}
   247   };
   248 
   249   /// \brief A controller with preset running time for the simulated annealing
   250   /// class.
   251   /// With this controller you can set the running time of the annealing
   252   /// process in advance. It works the following way: the controller measures
   253   /// a kind of divergence. The divergence is the difference of the average
   254   /// cost of the recently found solutions the cost of the best found one. In
   255   /// case this divergence is greater than a given threshold, then we decrease
   256   /// the annealing factor, that is we cool the system faster. In case the
   257   /// divergence is lower than the threshold, then we increase the temperature.
   258   /// The threshold is a function of the elapsed time which reaches zero at the
   259   /// desired end time.
   260   class AdvancedController : public ControllerBase {
   261   private:
   262     /// \brief Timer class to measure the elapsed time.
   263     Timer timer;
   264     /// \brief Calculates the threshold value.
   265     /// \param time the elapsed time in seconds
   266     virtual double threshold(double time) {
   267       return (-1.0) * start_threshold / end_time * time + start_threshold;
   268     }
   269     /// \brief Parameter used to calculate the running average.
   270     double alpha;
   271     /// \brief Parameter used to decrease the annealing factor.
   272     double beta;
   273     /// \brief Parameter used to increase the temperature.
   274     double gamma;
   275     /// \brief The time at the end of the algorithm.
   276     double end_time;
   277     /// \brief The time at the start of the algorithm.
   278     double start_time;
   279     /// \brief Starting threshold.
   280     double start_threshold;
   281     /// \brief Average cost of recent solutions.
   282     double avg_cost;
   283     /// \brief Temperature.
   284     double temp;
   285     /// \brief Annealing factor.
   286     double ann_fact;
   287     /// \brief Initial annealing factor.
   288     double init_ann_fact;
   289     /// \brief True when the annealing process has been started.
   290     bool start;
   291   public:
   292     /// \brief Constructor.
   293     /// \param _end_time running time in seconds
   294     /// \param _alpha parameter used to calculate the running average
   295     /// \param _beta parameter used to decrease the annealing factor
   296     /// \param _gamma parameter used to increase the temperature
   297     /// \param _ann_fact initial annealing factor
   298     AdvancedController(double _end_time, double _alpha = 0.2,
   299     double _beta = 0.9, double _gamma = 1.6, double _ann_fact = 0.9999) :
   300     alpha(_alpha), beta(_beta), gamma(_gamma), end_time(_end_time),
   301     ann_fact(_ann_fact), init_ann_fact(_ann_fact), start(false)
   302     {
   303       srand48(time(0));
   304     }
   305     /// \brief Does initializations before each run.
   306     void init() {
   307       avg_cost = simann->getCurrCost();
   308     }
   309     /// \brief This is called when a neighbouring state gets accepted.
   310     void acceptEvent() {
   311       avg_cost = alpha * simann->getCurrCost() + (1.0 - alpha) * avg_cost;
   312       if (!start) {
   313         static int cnt = 0;
   314         cnt++;
   315         if (cnt >= 100) {
   316           // calculate starting threshold and starting temperature
   317           start_threshold = 5.0 * fabs(simann->getBestCost() - avg_cost);
   318           temp = 10000.0;
   319           start = true;
   320           timer.restart();
   321         }
   322       }
   323     }
   324     /// \brief Decides whether to continue the annealing process or not.
   325     bool next() {
   326       if (!start) {
   327         return true;
   328       }
   329       else {
   330         double elapsed_time = timer.realTime();
   331         if (fabs(avg_cost - simann->getBestCost()) > threshold(elapsed_time)) {
   332           // decrease the annealing factor
   333           ann_fact *= beta;
   334         }
   335         else {
   336           // increase the temperature
   337           temp *= gamma;
   338           // reset the annealing factor
   339           ann_fact = init_ann_fact;
   340         }
   341         temp *= ann_fact;
   342         return elapsed_time < end_time;
   343       }
   344     }
   345     /// \brief Decides whether to accept the current solution or not.
   346     bool accept() {
   347       if (!start) {
   348         return true;
   349       }
   350       else {
   351         double cost_diff = simann->getCurrCost() - simann->getPrevCost();
   352         return (drand48() <= exp(-(cost_diff / temp)));
   353       }
   354     }
   355     /// \brief Destructor.
   356     virtual ~AdvancedController() {}
   357   };
   358 
   359 /// @}
   360 
   361 }
   362 
   363 #endif