lemon/simann.h
author deba
Fri, 27 Jan 2006 14:18:11 +0000
changeset 1915 f1f523d39d32
parent 1633 4bc163d55528
child 1918 09415ae11103
permissions -rw-r--r--
Add new ItemSetTraits for ANode and BNode
     1 #ifndef LEMON_SIMANN_H
     2 #define LEMON_SIMANN_H
     3 
     4 /// \ingroup experimental
     5 /// \file
     6 /// \brief Simulated annealing framework.
     7 ///
     8 /// \todo A test and some demo should be added
     9 /// \todo Doc should be improved
    10 /// \author Akos Ladanyi
    11 
    12 #include <cstdlib>
    13 #include <cmath>
    14 #include <lemon/time_measure.h>
    15 
    16 namespace lemon {
    17 
    18 /// \addtogroup experimental
    19 /// @{
    20 
    21   /*! \brief A base class for controllers. */
    22   class ControllerBase {
    23     friend class SimAnnBase;
    24   public:
    25     /*! \brief Pointer to the simulated annealing base class. */
    26     SimAnnBase *simann;
    27     /*! \brief Initializes the controller. */
    28     virtual void init() {}
    29     /*! \brief This is called when a neighbouring state gets accepted. */
    30     virtual void acceptEvent() {}
    31     /*! \brief This is called when the accepted neighbouring state's cost is
    32      *  less than the best found one's.
    33      */
    34     virtual void improveEvent() {}
    35     /*! \brief This is called when a neighbouring state gets rejected. */
    36     virtual void rejectEvent() {}
    37     /*! \brief Decides whether to continue the annealing process or not. */
    38     virtual bool next() = 0;
    39     /*! \brief Decides whether to accept the current solution or not. */
    40     virtual bool accept() = 0;
    41   };
    42 
    43   /*! \brief Skeleton of an entity class. */
    44   class EntityBase {
    45   public:
    46     /*! \brief Makes a minor change to the entity.
    47      *  \return the new cost
    48      */
    49     virtual double mutate() = 0;
    50     /*! \brief Restores the entity to its previous state i.e. reverts the
    51      *  effects of the last mutate().
    52      */
    53     virtual void revert() = 0;
    54     /*! \brief Makes a copy of the entity. */
    55     virtual EntityBase* clone() = 0;
    56     /*! \brief Makes a major change to the entity. */
    57     virtual void randomize() = 0;
    58   };
    59 
    60   /*! \brief Simulated annealing base class. */
    61   class SimAnnBase {
    62   private:
    63     /*! Pointer to the controller. */
    64     ControllerBase *controller;
    65     /*! \brief Cost of the current solution. */
    66     double curr_cost;
    67     /*! \brief Cost of the best solution. */
    68     double best_cost;
    69     /*! \brief Cost of the previous solution. */
    70     double prev_cost;
    71     /*! \brief Cost of the solution preceding the previous one. */
    72     double prev_prev_cost;
    73     /*! \brief Number of iterations. */
    74     long iter;
    75     /*! \brief Number of iterations which did not improve the solution since
    76      *  the last improvement. */
    77     long last_impr;
    78   protected:
    79     /*! \brief Step to a neighbouring state. */
    80     virtual double mutate() = 0;
    81     /*! \brief Reverts the last mutate(). */
    82     virtual void revert() = 0;
    83     /*! \brief Saves the current solution as the best one. */
    84     virtual void saveAsBest() = 0;
    85     /*! \brief Does initializations before each run. */
    86     virtual void init() {
    87       controller->init();
    88       curr_cost = prev_cost = prev_prev_cost = best_cost =
    89         std::numeric_limits<double>::infinity();
    90       iter = last_impr = 0;
    91     }
    92   public:
    93     /*! \brief Sets the controller class to use. */
    94     void setController(ControllerBase &_controller) {
    95       controller = &_controller;
    96       controller->simann = this;
    97     }
    98     /*! \brief Returns the cost of the current solution. */
    99     double getCurrCost() const { return curr_cost; }
   100     /*! \brief Returns the cost of the previous solution. */
   101     double getPrevCost() const { return prev_cost; }
   102     /*! \brief Returns the cost of the best solution. */
   103     double getBestCost() const { return best_cost; }
   104     /*! \brief Returns the number of iterations. */
   105     long getIter() const { return iter; }
   106     /*! \brief Returns the number of the last iteration when the solution was
   107      *  improved.
   108      */
   109     long getLastImpr() const { return last_impr; }
   110     /*! \brief Performs one iteration. */
   111     bool step() {
   112       iter++;
   113       prev_prev_cost = prev_cost;
   114       prev_cost = curr_cost;
   115       curr_cost = mutate();
   116       if (controller->accept()) {
   117         controller->acceptEvent();
   118         last_impr = iter;
   119         if (curr_cost < best_cost) {
   120           best_cost = curr_cost;
   121           saveAsBest();
   122           controller->improveEvent();
   123         }
   124       }
   125       else {
   126         revert();
   127         curr_cost = prev_cost;
   128         prev_cost = prev_prev_cost;
   129         controller->rejectEvent();
   130       }
   131       return controller->next();
   132     }
   133     /*! \brief Performs a given number of iterations.
   134      *  \param n the number of iterations
   135      */
   136     bool step(int n) {
   137       for(; n > 0 && step(); --n) ;
   138       return !n;
   139     }
   140     /*! \brief Starts the annealing process. */
   141     void run() {
   142       init();
   143       do { } while (step());
   144     }
   145   };
   146 
   147   /*! \brief Simulated annealing class. */
   148   class SimAnn : public SimAnnBase {
   149   private:
   150     /*! \brief Pointer to the current entity. */
   151     EntityBase *curr_ent;
   152     /*! \brief Pointer to the best entity. */
   153     EntityBase *best_ent;
   154     /*! \brief Does initializations before each run. */
   155     void init() {
   156       SimAnnBase::init();
   157       if (best_ent) delete best_ent;
   158       best_ent = NULL;
   159       curr_ent->randomize();
   160     }
   161   public:
   162     /*! \brief Constructor. */
   163     SimAnn() : curr_ent(NULL), best_ent(NULL) {}
   164     /*! \brief Destructor. */
   165     virtual ~SimAnn() {
   166       if (best_ent) delete best_ent;
   167     }
   168     /*! \brief Step to a neighbouring state. */
   169     double mutate() {
   170       return curr_ent->mutate();
   171     }
   172     /*! \brief Reverts the last mutate(). */
   173     void revert() {
   174       curr_ent->revert();
   175     }
   176     /*! \brief Saves the current solution as the best one. */
   177     void saveAsBest() { 
   178       if (best_ent) delete best_ent;
   179       best_ent = curr_ent->clone();
   180     }
   181     /*! \brief Sets the current entity. */
   182     void setEntity(EntityBase &_ent) {
   183       curr_ent = &_ent;
   184     }
   185     /*! \brief Returns a copy of the best found entity. */
   186     EntityBase* getBestEntity() { return best_ent->clone(); }
   187   };
   188 
   189   /*! \brief A simple controller for the simulated annealing class. */
   190   class SimpleController : public ControllerBase {
   191   public:
   192     /*! \brief Maximum number of iterations. */
   193     long max_iter;
   194     /*! \brief Maximum number of iterations which do not improve the
   195      *  solution. */
   196     long max_no_impr;
   197     /*! \brief Temperature. */
   198     double temp;
   199     /*! \brief Annealing factor. */
   200     double ann_fact;
   201     /*! \brief Constructor.
   202      *  \param _max_iter maximum number of iterations
   203      *  \param _max_no_impr maximum number of consecutive iterations which do
   204      *         not yield a better solution
   205      *  \param _temp initial temperature
   206      *  \param _ann_fact annealing factor
   207      */
   208     SimpleController(long _max_iter = 500000, long _max_no_impr = 20000,
   209     double _temp = 1000.0, double _ann_fact = 0.9999) : max_iter(_max_iter),
   210       max_no_impr(_max_no_impr), temp(_temp), ann_fact(_ann_fact)
   211     {
   212       srand48(time(0));
   213     }
   214     /*! \brief This is called when a neighbouring state gets accepted. */
   215     void acceptEvent() {}
   216     /*! \brief This is called when the accepted neighbouring state's cost is
   217      *  less than the best found one's.
   218      */
   219     void improveEvent() {}
   220     /*! \brief This is called when a neighbouring state gets rejected. */
   221     void rejectEvent() {}
   222     /*! \brief Decides whether to continue the annealing process or not. Also
   223      *  decreases the temperature. */
   224     bool next() {
   225       temp *= ann_fact;
   226       bool quit = (simann->getIter() > max_iter) ||
   227         (simann->getIter() - simann->getLastImpr() > max_no_impr);
   228       return !quit;
   229     }
   230     /*! \brief Decides whether to accept the current solution or not. */
   231     bool accept() {
   232       double cost_diff = simann->getPrevCost() - simann->getCurrCost();
   233       return (drand48() <= exp(cost_diff / temp));
   234     }
   235   };
   236 
   237   /*! \brief A controller with preset running time for the simulated annealing
   238    *  class.
   239    *
   240    *  With this controller you can set the running time of the annealing
   241    *  process in advance. It works the following way: the controller measures
   242    *  a kind of divergence. The divergence is the difference of the average
   243    *  cost of the recently found solutions the cost of the best found one. In
   244    *  case this divergence is greater than a given threshold, then we decrease
   245    *  the annealing factor, that is we cool the system faster. In case the
   246    *  divergence is lower than the threshold, then we increase the temperature.
   247    *  The threshold is a function of the elapsed time which reaches zero at the
   248    *  desired end time.
   249    */
   250   class AdvancedController : public ControllerBase {
   251   private:
   252     Timer timer;
   253     /*! \param time the elapsed time in seconds */
   254     virtual double threshold(double time) {
   255       return (-1.0) * start_threshold / end_time * time + start_threshold;
   256     }
   257   public:
   258     double alpha;
   259     double beta;
   260     double gamma;
   261     /*! \brief The time at the end of the algorithm. */
   262     double end_time;
   263     /*! \brief The time at the start of the algorithm. */
   264     double start_time;
   265     /*! \brief Starting threshold. */
   266     double start_threshold;
   267     /*! \brief Average cost of recent solutions. */
   268     double avg_cost;
   269     /*! \brief Temperature. */
   270     double temp;
   271     /*! \brief Annealing factor. */
   272     double ann_fact;
   273     /*! \brief Initial annealing factor. */
   274     double init_ann_fact;
   275     bool warmup;
   276     /*! \brief Constructor.
   277      *  \param _end_time running time in seconds
   278      *  \param _alpha parameter used to calculate the running average
   279      *  \param _beta parameter used to decrease the annealing factor
   280      *  \param _gamma parameter used to increase the temperature
   281      *  \param _ann_fact initial annealing factor
   282      */
   283     AdvancedController(double _end_time, double _alpha = 0.2,
   284     double _beta = 0.9, double _gamma = 1.6, double _ann_fact = 0.9999) :
   285     alpha(_alpha), beta(_beta), gamma(_gamma), end_time(_end_time),
   286     ann_fact(_ann_fact), init_ann_fact(_ann_fact), warmup(true)
   287     {
   288       srand48(time(0));
   289     }
   290     void init() {
   291       avg_cost = simann->getCurrCost();
   292     }
   293     /*! \brief This is called when a neighbouring state gets accepted. */
   294     void acceptEvent() {
   295       avg_cost = alpha * simann->getCurrCost() + (1.0 - alpha) * avg_cost;
   296       if (warmup) {
   297         static int cnt = 0;
   298         cnt++;
   299         if (cnt >= 100) {
   300           // calculate starting threshold and starting temperature
   301           start_threshold = 5.0 * fabs(simann->getBestCost() - avg_cost);
   302           temp = 10000.0;
   303           warmup = false;
   304           timer.restart();
   305         }
   306       }
   307     }
   308     /*! \brief Decides whether to continue the annealing process or not. */
   309     bool next() {
   310       if (warmup) {
   311         return true;
   312       }
   313       else {
   314         double elapsed_time = timer.getRealTime();
   315         if (fabs(avg_cost - simann->getBestCost()) > threshold(elapsed_time)) {
   316           // decrease the annealing factor
   317           ann_fact *= beta;
   318         }
   319         else {
   320           // increase the temperature
   321           temp *= gamma;
   322           // reset the annealing factor
   323           ann_fact = init_ann_fact;
   324         }
   325         temp *= ann_fact;
   326         return elapsed_time < end_time;
   327       }
   328     }
   329     /*! \brief Decides whether to accept the current solution or not. */
   330     bool accept() {
   331       if (warmup) {
   332         // we accept eveything during the "warm up" phase
   333         return true;
   334       }
   335       else {
   336         double cost_diff = simann->getPrevCost() - simann->getCurrCost();
   337         if (cost_diff < 0.0) {
   338           return (drand48() <= exp(cost_diff / temp));
   339         }
   340         else {
   341           return true;
   342         }
   343       }
   344     }
   345   };
   346 
   347 /// @}
   348 
   349 }
   350 
   351 #endif