COIN-OR::LEMON - Graph Library

source: lemon-0.x/lemon/simann.h @ 2229:4dbb6dd2dd4b

Last change on this file since 2229:4dbb6dd2dd4b was 2229:4dbb6dd2dd4b, checked in by Balazs Dezso, 17 years ago

Mersenne Twister random number generator

The code is based on the official MT19937 implementation
It is fully rewritten:

http://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/emt.html

todo: fixing copyright information

File size: 12.5 KB
Line 
1/* -*- C++ -*-
2 *
3 * This file is a part of LEMON, a generic C++ optimization library
4 *
5 * Copyright (C) 2003-2006
6 * Egervary Jeno Kombinatorikus Optimalizalasi Kutatocsoport
7 * (Egervary Research Group on Combinatorial Optimization, EGRES).
8 *
9 * Permission to use, modify and distribute this software is granted
10 * provided that this copyright notice appears in all copies. For
11 * precise terms see the accompanying LICENSE file.
12 *
13 * This software is provided "AS IS" with no warranty of any kind,
14 * express or implied, and with no claim as to its suitability for any
15 * purpose.
16 *
17 */
18
19#ifndef LEMON_SIMANN_H
20#define LEMON_SIMANN_H
21
22/// \ingroup experimental
23/// \file
24/// \brief Simulated annealing framework.
25///
26/// \todo A test and some demo should be added
27/// \todo Doc should be improved
28/// \author Akos Ladanyi
29
30#include <cstdlib>
31#include <cmath>
32#include <limits>
33#include <lemon/time_measure.h>
34#include <lemon/random.h>
35
36namespace lemon {
37
38/// \addtogroup experimental
39/// @{
40
41  class SimAnnBase;
42
43  /// \brief A base class for controllers.
44  class ControllerBase {
45  public:
46    friend class SimAnnBase;
47    /// \brief Pointer to the simulated annealing base class.
48    SimAnnBase *simann;
49    /// \brief Initializes the controller.
50    virtual void init() {}
51    /// \brief This is called by the simulated annealing class when a
52    /// neighbouring state gets accepted.
53    virtual void acceptEvent() {}
54    /// \brief This is called by the simulated annealing class when the
55    /// accepted neighbouring state's cost is less than the best found one's.
56    virtual void improveEvent() {}
57    /// \brief This is called by the simulated annealing class when a
58    /// neighbouring state gets rejected.
59    virtual void rejectEvent() {}
60    /// \brief Decides whether to continue the annealing process or not.
61    virtual bool next() = 0;
62    /// \brief Decides whether to accept the current solution or not.
63    virtual bool accept() = 0;
64    /// \brief Destructor.
65    virtual ~ControllerBase() {}
66  };
67
68  /// \brief Skeleton of an entity class.
69  class EntityBase {
70  public:
71    /// \brief Makes a minor change to the entity.
72    /// \return the new cost
73    virtual double mutate() = 0;
74    /// \brief Restores the entity to its previous state i.e. reverts the
75    /// effects of the last mutate().
76    virtual void revert() = 0;
77    /// \brief Makes a copy of the entity.
78    virtual EntityBase* clone() = 0;
79    /// \brief Makes a major change to the entity.
80    virtual void randomize() = 0;
81    /// \brief Destructor.
82    virtual ~EntityBase() {}
83  };
84
85  /// \brief Simulated annealing abstract base class.
86  /// Can be used to derive a custom simulated annealing class if \ref SimAnn
87  /// doesn't fit your needs.
88  class SimAnnBase {
89  private:
90    /// \brief Pointer to the controller.
91    ControllerBase *controller;
92    /// \brief Cost of the current solution.
93    double curr_cost;
94    /// \brief Cost of the best solution.
95    double best_cost;
96    /// \brief Cost of the previous solution.
97    double prev_cost;
98    /// \brief Cost of the solution preceding the previous one.
99    double prev_prev_cost;
100    /// \brief Number of iterations.
101    long iter;
102    /// \brief Number of iterations which did not improve the solution since
103    /// the last improvement.
104    long last_impr;
105  protected:
106    /// \brief Step to a neighbouring state.
107    virtual double mutate() = 0;
108    /// \brief Reverts the last mutate().
109    virtual void revert() = 0;
110    /// \brief Saves the current solution as the best one.
111    virtual void saveAsBest() = 0;
112    /// \brief Does initializations before each run.
113    virtual void init() {
114      controller->init();
115      curr_cost = prev_cost = prev_prev_cost = best_cost =
116        std::numeric_limits<double>::infinity();
117      iter = last_impr = 0;
118    }
119  public:
120    /// \brief Sets the controller class to use.
121    void setController(ControllerBase &_controller) {
122      controller = &_controller;
123      controller->simann = this;
124    }
125    /// \brief Returns the cost of the current solution.
126    double getCurrCost() const { return curr_cost; }
127    /// \brief Returns the cost of the previous solution.
128    double getPrevCost() const { return prev_cost; }
129    /// \brief Returns the cost of the best solution.
130    double getBestCost() const { return best_cost; }
131    /// \brief Returns the number of iterations done.
132    long getIter() const { return iter; }
133    /// \brief Returns the ordinal number of the last iteration when the
134    /// solution was improved.
135    long getLastImpr() const { return last_impr; }
136    /// \brief Performs one iteration.
137    bool step() {
138      iter++;
139      prev_prev_cost = prev_cost;
140      prev_cost = curr_cost;
141      curr_cost = mutate();
142      if (controller->accept()) {
143        controller->acceptEvent();
144        last_impr = iter;
145        if (curr_cost < best_cost) {
146          best_cost = curr_cost;
147          saveAsBest();
148          controller->improveEvent();
149        }
150      }
151      else {
152        revert();
153        curr_cost = prev_cost;
154        prev_cost = prev_prev_cost;
155        controller->rejectEvent();
156      }
157      return controller->next();
158    }
159    /// \brief Performs a given number of iterations.
160    /// \param n the number of iterations
161    bool step(int n) {
162      for(; n > 0 && step(); --n) ;
163      return !n;
164    }
165    /// \brief Starts the annealing process.
166    void run() {
167      init();
168      do { } while (step());
169    }
170    /// \brief Destructor.
171    virtual ~SimAnnBase() {}
172  };
173
174  /// \brief Simulated annealing class.
175  class SimAnn : public SimAnnBase {
176  private:
177    /// \brief Pointer to the current entity.
178    EntityBase *curr_ent;
179    /// \brief Pointer to the best entity.
180    EntityBase *best_ent;
181    /// \brief Does initializations before each run.
182    void init() {
183      SimAnnBase::init();
184      if (best_ent) delete best_ent;
185      best_ent = NULL;
186      curr_ent->randomize();
187    }
188  public:
189    /// \brief Constructor.
190    SimAnn() : curr_ent(NULL), best_ent(NULL) {}
191    /// \brief Destructor.
192    virtual ~SimAnn() {
193      if (best_ent) delete best_ent;
194    }
195    /// \brief Step to a neighbouring state.
196    double mutate() {
197      return curr_ent->mutate();
198    }
199    /// \brief Reverts the last mutate().
200    void revert() {
201      curr_ent->revert();
202    }
203    /// \brief Saves the current solution as the best one.
204    void saveAsBest() {
205      if (best_ent) delete best_ent;
206      best_ent = curr_ent->clone();
207    }
208    /// \brief Sets the current entity.
209    void setEntity(EntityBase &_ent) {
210      curr_ent = &_ent;
211    }
212    /// \brief Returns a copy of the best found entity.
213    EntityBase* getBestEntity() { return best_ent->clone(); }
214  };
215
216  /// \brief A simple controller for the simulated annealing class.
217  /// This controller starts from a given initial temperature and evenly
218  /// decreases it.
219  class SimpleController : public ControllerBase {
220  private:
221    /// \brief Maximum number of iterations.
222    long max_iter;
223    /// \brief Maximum number of iterations which do not improve the
224    /// solution.
225    long max_no_impr;
226    /// \brief Temperature.
227    double temp;
228    /// \brief Annealing factor.
229    double ann_fact;
230    /// \brief Constructor.
231    /// \param _max_iter maximum number of iterations
232    /// \param _max_no_impr maximum number of consecutive iterations which do
233    ///        not yield a better solution
234    /// \param _temp initial temperature
235    /// \param _ann_fact annealing factor
236  public:
237    SimpleController(long _max_iter = 500000, long _max_no_impr = 20000,
238    double _temp = 1000.0, double _ann_fact = 0.9999) : max_iter(_max_iter),
239      max_no_impr(_max_no_impr), temp(_temp), ann_fact(_ann_fact)
240    {
241    }
242    /// \brief This is called when a neighbouring state gets accepted.
243    void acceptEvent() {}
244    /// \brief This is called when the accepted neighbouring state's cost is
245    /// less than the best found one's.
246    void improveEvent() {}
247    /// \brief This is called when a neighbouring state gets rejected.
248    void rejectEvent() {}
249    /// \brief Decides whether to continue the annealing process or not. Also
250    /// decreases the temperature.
251    bool next() {
252      temp *= ann_fact;
253      bool quit = (simann->getIter() > max_iter) ||
254        (simann->getIter() - simann->getLastImpr() > max_no_impr);
255      return !quit;
256    }
257    /// \brief Decides whether to accept the current solution or not.
258    bool accept() {
259      double cost_diff = simann->getCurrCost() - simann->getPrevCost();
260      return (rnd.getReal() <= exp(-(cost_diff / temp)));
261    }
262    /// \brief Destructor.
263    virtual ~SimpleController() {}
264  };
265
266  /// \brief A controller with preset running time for the simulated annealing
267  /// class.
268  /// With this controller you can set the running time of the annealing
269  /// process in advance. It works the following way: the controller measures
270  /// a kind of divergence. The divergence is the difference of the average
271  /// cost of the recently found solutions the cost of the best found one. In
272  /// case this divergence is greater than a given threshold, then we decrease
273  /// the annealing factor, that is we cool the system faster. In case the
274  /// divergence is lower than the threshold, then we increase the temperature.
275  /// The threshold is a function of the elapsed time which reaches zero at the
276  /// desired end time.
277  class AdvancedController : public ControllerBase {
278  private:
279    /// \brief Timer class to measure the elapsed time.
280    Timer timer;
281    /// \brief Calculates the threshold value.
282    /// \param time the elapsed time in seconds
283    virtual double threshold(double time) {
284      return (-1.0) * start_threshold / end_time * time + start_threshold;
285    }
286    /// \brief Parameter used to calculate the running average.
287    double alpha;
288    /// \brief Parameter used to decrease the annealing factor.
289    double beta;
290    /// \brief Parameter used to increase the temperature.
291    double gamma;
292    /// \brief The time at the end of the algorithm.
293    double end_time;
294    /// \brief The time at the start of the algorithm.
295    double start_time;
296    /// \brief Starting threshold.
297    double start_threshold;
298    /// \brief Average cost of recent solutions.
299    double avg_cost;
300    /// \brief Temperature.
301    double temp;
302    /// \brief Annealing factor.
303    double ann_fact;
304    /// \brief Initial annealing factor.
305    double init_ann_fact;
306    /// \brief True when the annealing process has been started.
307    bool start;
308  public:
309    /// \brief Constructor.
310    /// \param _end_time running time in seconds
311    /// \param _alpha parameter used to calculate the running average
312    /// \param _beta parameter used to decrease the annealing factor
313    /// \param _gamma parameter used to increase the temperature
314    /// \param _ann_fact initial annealing factor
315    AdvancedController(double _end_time, double _alpha = 0.2,
316    double _beta = 0.9, double _gamma = 1.6, double _ann_fact = 0.9999) :
317    alpha(_alpha), beta(_beta), gamma(_gamma), end_time(_end_time),
318    ann_fact(_ann_fact), init_ann_fact(_ann_fact), start(false)
319    {
320    }
321    /// \brief Does initializations before each run.
322    void init() {
323      avg_cost = simann->getCurrCost();
324    }
325    /// \brief This is called when a neighbouring state gets accepted.
326    void acceptEvent() {
327      avg_cost = alpha * simann->getCurrCost() + (1.0 - alpha) * avg_cost;
328      if (!start) {
329        static int cnt = 0;
330        cnt++;
331        if (cnt >= 100) {
332          // calculate starting threshold and starting temperature
333          start_threshold = 5.0 * fabs(simann->getBestCost() - avg_cost);
334          temp = 10000.0;
335          start = true;
336          timer.restart();
337        }
338      }
339    }
340    /// \brief Decides whether to continue the annealing process or not.
341    bool next() {
342      if (!start) {
343        return true;
344      }
345      else {
346        double elapsed_time = timer.realTime();
347        if (fabs(avg_cost - simann->getBestCost()) > threshold(elapsed_time)) {
348          // decrease the annealing factor
349          ann_fact *= beta;
350        }
351        else {
352          // increase the temperature
353          temp *= gamma;
354          // reset the annealing factor
355          ann_fact = init_ann_fact;
356        }
357        temp *= ann_fact;
358        return elapsed_time < end_time;
359      }
360    }
361    /// \brief Decides whether to accept the current solution or not.
362    bool accept() {
363      if (!start) {
364        return true;
365      }
366      else {
367        double cost_diff = simann->getCurrCost() - simann->getPrevCost();
368        return (rnd.getReal() <= exp(-(cost_diff / temp)));
369      }
370    }
371    /// \brief Destructor.
372    virtual ~AdvancedController() {}
373  };
374
375/// @}
376
377}
378
379#endif
Note: See TracBrowser for help on using the repository browser.