rgsl/types/
siman.rs

1//
2// A rust binding for the GSL library by Guillaume Gomez (guillaume1.gomez@gmail.com)
3//
4
5/*!
6# 25 Simulated Annealing
7
8Stochastic search techniques are used when the structure of a space is not well understood or
9is not smooth, so that techniques like Newton’s method (which requires calculating Jacobian
10derivative matrices) cannot be used. In particular, these techniques are frequently used to
11solve combinatorial optimization problems, such as the traveling salesman problem.
12
13The goal is to find a point in the space at which a real valued energy function (or cost
14function) is minimized. Simulated annealing is a minimization technique which has given
15good results in avoiding local minima; it is based on the idea of taking a random walk
16through the space at successively lower temperatures, where the probability of taking a
17step is given by a Boltzmann distribution.
18
19The functions described in this chapter are declared in the header file gsl_siman.h.
20
21## Simulated Annealing algorithm
22
23The simulated annealing algorithm takes random walks through the problem space, looking
24for points with low energies; in these random walks, the probability of taking a step is
25determined by the Boltzmann distribution,
26
27> p = e −(E i+1 −E i )/(kT )
28> if E i+1 > E i , and p = 1 when E i+1 ≤ E i
29
30In other words, a step will occur if the new energy is lower. If the new energy is higher,
31the transition can still occur, and its likelihood is proportional to the temperature T and
32inversely proportional to the energy difference E i+1 − E i .
33
34The temperature T is initially set to a high value, and a random walk is carried out
35at that temperature. Then the temperature is lowered very slightly according to a cooling
36schedule, for example: T → T /μ T where μ T is slightly greater than 1.
37
38The slight probability of taking a step that gives higher energy is what allows simulated
39annealing to frequently get out of local minima.
40!*/
41
42const GSL_LOG_DBL_MIN: f64 = -7.0839641853226408e+02;
43
44pub struct SimAnnealing<T: Clone> {
45    x0_p: T,
46    params: SimAnnealingParams,
47    Efunc_t: gsl_siman_Efunc_t<T>,
48    step_t: gsl_siman_step_t<T>,
49    metric_t: gsl_siman_metric_t<T>,
50    print_t: Option<gsl_siman_print_t<T>>,
51}
52
53type gsl_siman_Efunc_t<T> = fn(&T) -> f64;
54type gsl_siman_step_t<T> = fn(&mut crate::Rng, &mut T, f64);
55type gsl_siman_metric_t<T> = fn(&T, &T) -> f64;
56type gsl_siman_print_t<T> = fn(&T);
57
58fn boltzmann(E: f64, new_E: f64, T: f64, params: &SimAnnealingParams) -> f64 {
59    let x = -(new_E - E) / (params.k * T);
60    // avoid underflow errors for large uphill steps
61    if x < GSL_LOG_DBL_MIN {
62        0.0
63    } else {
64        x.exp()
65    }
66}
67
68impl<T> SimAnnealing<T>
69where
70    T: Clone,
71{
72    pub fn new(
73        x0_p: T,
74        ef: gsl_siman_Efunc_t<T>,
75        take_step: gsl_siman_step_t<T>,
76        distance: gsl_siman_metric_t<T>,
77        print_pos: Option<gsl_siman_print_t<T>>,
78        params: SimAnnealingParams,
79    ) -> SimAnnealing<T> {
80        SimAnnealing {
81            x0_p,
82            params,
83            Efunc_t: ef,
84            step_t: take_step,
85            metric_t: distance,
86            print_t: print_pos,
87        }
88    }
89
90    /*
91    /* implementation of a basic simulated annealing algorithm */
92
93    void
94    gsl_siman_solve (const gsl_rng * r, void *x0_p, gsl_siman_Efunc_t Ef,
95                    gsl_siman_step_t take_step,
96                    gsl_siman_metric_t distance,
97                    gsl_siman_print_t print_position,
98                    gsl_siman_copy_t copyfunc,
99                    gsl_siman_copy_construct_t copy_constructor,
100                    gsl_siman_destroy_t destructor,
101                    size_t element_size,
102                    gsl_siman_params_t params)
103    {
104        void *x, *new_x, *best_x;
105        double E, new_E, best_E;
106        int i;
107        double T, T_factor;
108        int n_evals = 1, n_iter = 0, n_accepts, n_rejects, n_eless;
109
110        /* this function requires that either the dynamic functions (copy,
111            copy_constructor and destrcutor) are passed, or that an element
112            size is given */
113        assert((copyfunc != NULL && copy_constructor != NULL && destructor != NULL)
114                || (element_size != 0));
115
116        distance = 0 ; /* This parameter is not currently used */
117        E = Ef(x0_p);
118
119        if (copyfunc) {
120            x = copy_constructor(x0_p);
121            new_x = copy_constructor(x0_p);
122            best_x = copy_constructor(x0_p);
123        } else {
124            x = (void *) malloc (element_size);
125            memcpy (x, x0_p, element_size);
126            new_x = (void *) malloc (element_size);
127            best_x =  (void *) malloc (element_size);
128            memcpy (best_x, x0_p, element_size);
129        }
130
131        best_E = E;
132
133        T = params.t_initial;
134        T_factor = 1.0 / params.mu_t;
135
136        if (print_position) {
137            printf ("#-iter  #-evals   temperature     position   energy\n");
138        }
139
140        while (1) {
141
142            n_accepts = 0;
143            n_rejects = 0;
144            n_eless = 0;
145
146            for (i = 0; i < params.iters_fixed_T; ++i) {
147
148                copy_state(x, new_x, element_size, copyfunc);
149
150                take_step (r, new_x, params.step_size);
151                new_E = Ef (new_x);
152
153                if(new_E <= best_E){
154                    if (copyfunc) {
155                        copyfunc(new_x,best_x);
156                    } else {
157                        memcpy (best_x, new_x, element_size);
158                    }
159                    best_E=new_E;
160                }
161
162                ++n_evals;                /* keep track of Ef() evaluations */
163                /* now take the crucial step: see if the new point is accepted
164                    or not, as determined by the boltzmann probability */
165                if (new_E < E) {
166                    if (new_E < best_E) {
167                        copy_state(new_x, best_x, element_size, copyfunc);
168                        best_E = new_E;
169                    }
170
171                    /* yay! take a step */
172                    copy_state(new_x, x, element_size, copyfunc);
173                    E = new_E;
174                    ++n_eless;
175
176                } else if (gsl_rng_uniform(r) < boltzmann(E, new_E, T, &params)) {
177                    /* yay! take a step */
178                    copy_state(new_x, x, element_size, copyfunc);
179                        E = new_E;
180                        ++n_accepts;
181
182                } else {
183                    ++n_rejects;
184                }
185            }
186
187            if (print_position) {
188                /* see if we need to print stuff as we go */
189                /*       printf("%5d %12g %5d %3d %3d %3d", n_iter, T, n_evals, */
190                /*           100*n_eless/n_steps, 100*n_accepts/n_steps, */
191                /*           100*n_rejects/n_steps); */
192                printf ("%5d   %7d  %12g", n_iter, n_evals, T);
193                print_position (x);
194                printf ("  %12g  %12g\n", E, best_E);
195            }
196
197            /* apply the cooling schedule to the temperature */
198            /* FIXME: I should also introduce a cooling schedule for the iters */
199            T *= T_factor;
200            ++n_iter;
201            if (T < params.t_min) {
202                break;
203            }
204        }
205
206        /* at the end, copy the result onto the initial point, so we pass it
207            back to the caller */
208        copy_state(best_x, x0_p, element_size, copyfunc);
209
210        if (copyfunc) {
211            destructor(x);
212            destructor(new_x);
213            destructor(best_x);
214        } else {
215            free (x);
216            free (new_x);
217            free (best_x);
218        }
219    }
220    */
221    /// This function performs a simulated annealing search through a given space. The space
222    /// is specified by providing the functions Ef and distance. The simulated annealing steps
223    /// are generated using the random number generator `rng` and the function `take_step`.
224    ///
225    /// The starting configuration of the system should be given by x0\_p.
226    ///
227    /// The params structure (described below) controls the run by providing the temperature
228    /// schedule and other tunable parameters to the algorithm.
229    ///
230    /// On exit the best result achieved during the search is returned. If the annealing
231    /// process has been successful this should be a good approximation to the optimal point
232    /// in the space.
233    ///
234    /// If the argument `print_pos` is not None, a debugging log will be printed to
235    /// stdout with the following columns: ```#-iter #-evals temperature position energy best_energy```
236    /// and the output of the function print position itself.
237    pub fn solve(&self, rng: &mut crate::Rng) -> T {
238        let mut x = self.x0_p.clone();
239        let mut new_x = self.x0_p.clone();
240        let mut best_x = self.x0_p.clone();
241
242        let mut n_evals = 0_usize;
243
244        let mut E = (self.Efunc_t)(&self.x0_p);
245        let mut best_E = E;
246
247        let mut Temp = self.params.t_initial;
248        let Temp_factor = 1.0 / self.params.mu_t;
249
250        if self.print_t.is_some() {
251            println!(
252                "{i:^6} | {e:^7} | {t:^12} | {p:^15} | {E:^13}",
253                i = "#-iter",
254                e = "#-evals",
255                t = "temperature",
256                p = "position",
257                E = "energy"
258            );
259        }
260
261        let mut iter = 0;
262        loop {
263            //let mut n_accepts = 0;
264            //let mut n_rejects = 0;
265            //let mut n_eless = 0;
266
267            for _ in 0..self.params.iters_fixed_T {
268                x = new_x.clone();
269
270                (self.step_t)(rng, &mut new_x, self.params.step_size);
271                let new_E = (self.Efunc_t)(&new_x);
272                n_evals += 1; // keep track of Ef() evaluations
273
274                if new_E <= best_E {
275                    best_x = new_x.clone();
276                    best_E = new_E;
277                }
278
279                // now take the crucial step: see if the new point is accepted
280                // or not, as determined by the boltzmann probability
281                if new_E < E {
282                    if new_E < best_E {
283                        best_x = new_x.clone();
284                        best_E = new_E;
285                    }
286                    // yay! take a step
287                    x = new_x.clone();
288                    E = new_E;
289                //n_eless += 1;
290                } else if rng.uniform() < boltzmann(E, new_E, Temp, &self.params) {
291                    // yay! take a step
292                    x = new_x.clone();
293                    E = new_E;
294                    //n_accepts += 1;
295                }
296            }
297
298            if let Some(ref printf) = self.print_t {
299                print!("{:>06} | {:>07} | {:>12.10} | ", iter, n_evals, Temp);
300                printf(&x);
301                println!(" | {:+>13.12}", E);
302            }
303
304            Temp *= Temp_factor;
305            iter += 1;
306            if Temp < self.params.t_min {
307                break;
308            }
309        }
310
311        best_x
312    }
313
314    /*
315    /* implementation of a simulated annealing algorithm with many tries */
316
317    void
318    gsl_siman_solve_many (const gsl_rng * r, void *x0_p, gsl_siman_Efunc_t Ef,
319                        gsl_siman_step_t take_step,
320                        gsl_siman_metric_t distance,
321                        gsl_siman_print_t print_position,
322                        size_t element_size,
323                        gsl_siman_params_t params)
324    {
325        /* the new set of trial points, and their energies and probabilities */
326        void *x, *new_x;
327        double *energies, *probs, *sum_probs;
328        double Ex;                    /* energy of the chosen point */
329        double T, T_factor;           /* the temperature and a step multiplier */
330        int i;
331        double u;                     /* throw the die to choose a new "x" */
332        int n_iter;
333
334        if (print_position) {
335            printf ("#-iter    temperature       position");
336            printf ("         delta_pos        energy\n");
337        }
338
339        x = (void *) malloc (params.n_tries * element_size);
340        new_x = (void *) malloc (params.n_tries * element_size);
341        energies = (double *) malloc (params.n_tries * sizeof (double));
342        probs = (double *) malloc (params.n_tries * sizeof (double));
343        sum_probs = (double *) malloc (params.n_tries * sizeof (double));
344
345        T = params.t_initial;
346        T_factor = 1.0 / params.mu_t;
347
348        memcpy (x, x0_p, element_size);
349
350        n_iter = 0;
351        while (1)
352            {
353            Ex = Ef (x);
354            for (i = 0; i < params.n_tries - 1; ++i)
355                {                       /* only go to N_TRIES-2 */
356                /* center the new_x[] around x, then pass it to take_step() */
357                sum_probs[i] = 0;
358                memcpy ((char *)new_x + i * element_size, x, element_size);
359                take_step (r, (char *)new_x + i * element_size, params.step_size);
360                energies[i] = Ef ((char *)new_x + i * element_size);
361                probs[i] = boltzmann(Ex, energies[i], T, &params);
362                }
363            /* now add in the old value of "x", so it is a contendor */
364            memcpy ((char *)new_x + (params.n_tries - 1) * element_size, x, element_size);
365            energies[params.n_tries - 1] = Ex;
366            probs[params.n_tries - 1] = boltzmann(Ex, energies[i], T, &params);
367
368            /* now throw biased die to see which new_x[i] we choose */
369            sum_probs[0] = probs[0];
370            for (i = 1; i < params.n_tries; ++i)
371                {
372                sum_probs[i] = sum_probs[i - 1] + probs[i];
373                }
374            u = gsl_rng_uniform (r) * sum_probs[params.n_tries - 1];
375            for (i = 0; i < params.n_tries; ++i)
376                {
377                if (u < sum_probs[i])
378                    {
379                    memcpy (x, (char *) new_x + i * element_size, element_size);
380                    break;
381                    }
382                }
383            if (print_position)
384                {
385                printf ("%5d\t%12g\t", n_iter, T);
386                print_position (x);
387                printf ("\t%12g\t%12g\n", distance (x, x0_p), Ex);
388                }
389            T *= T_factor;
390            ++n_iter;
391            if (T < params.t_min)
392            {
393            break;
394                }
395            }
396
397        /* now return the value via x0_p */
398        memcpy (x0_p, x, element_size);
399
400        /*  printf("the result is: %g (E=%g)\n", x, Ex); */
401
402        free (x);
403        free (new_x);
404        free (energies);
405        free (probs);
406        free (sum_probs);
407    }
408    */
409    /// Like the function solve, but performs multiple runs and returns the best result.
410    pub fn solve_many(&self, rng: &mut crate::Rng) -> T {
411        let mut x = self.x0_p.clone();
412        let mut new_x = Vec::with_capacity(self.params.n_tries);
413
414        let mut energies = Vec::with_capacity(self.params.n_tries);
415        let mut probs = Vec::with_capacity(self.params.n_tries);
416        let mut sum_probs = Vec::with_capacity(self.params.n_tries);
417
418        let mut Temp = self.params.t_initial;
419        let Temp_factor = 1.0 / self.params.mu_t;
420
421        if self.print_t.is_some() {
422            println!(
423                "{i:^6} | {t:^12} | {p:^15} | {d:^15} | {E:^13}",
424                i = "#-iter",
425                t = "temperature",
426                p = "position",
427                d = "delta_pos",
428                E = "energy"
429            );
430        }
431
432        let mut iter = 0;
433        loop {
434            let Ex = (self.Efunc_t)(&x);
435            for i in 0..self.params.n_tries - 1 {
436                // only go to N_TRIES-2
437                sum_probs.push(0.0);
438                new_x.push(x.clone());
439
440                (self.step_t)(rng, &mut new_x[i], self.params.step_size);
441                energies.push((self.Efunc_t)(&new_x[i]));
442                probs.push(boltzmann(Ex, energies[i], Temp, &self.params));
443            }
444
445            // now add in the old value of "x", so it is a contendor
446            new_x.push(x.clone());
447            energies.push(Ex);
448            probs.push(boltzmann(
449                Ex,
450                energies[self.params.n_tries - 2],
451                Temp,
452                &self.params,
453            ));
454            sum_probs.push(0.0);
455
456            // now throw biased die to see which new_x[i] we choose
457            sum_probs[0] = probs[0];
458            for i in 1..self.params.n_tries {
459                sum_probs[i] = sum_probs[i - 1] + probs[i];
460            }
461            let u = rng.uniform() * *sum_probs.last().unwrap();
462            for i in 0..self.params.n_tries {
463                if u < sum_probs[i] {
464                    x = new_x[i].clone();
465                    break;
466                }
467            }
468
469            if let Some(ref printf) = self.print_t {
470                print!("{:>06} | {:>12.10} | ", iter, Temp);
471                printf(&x);
472                println!(
473                    " | {: >15.12} | {: >13.12}",
474                    (self.metric_t)(&x, &self.x0_p),
475                    Ex
476                );
477            }
478
479            Temp *= Temp_factor;
480            iter += 1;
481            if Temp < self.params.t_min {
482                break;
483            }
484        }
485        x
486    }
487}
488
489pub struct SimAnnealingParams {
490    n_tries: usize,
491    iters_fixed_T: usize,
492    step_size: f64,
493    k: f64,
494    t_initial: f64,
495    mu_t: f64,
496    t_min: f64,
497}
498
499impl SimAnnealingParams {
500    /// These are the parameters that control a run of the simulated annealing algorithm.
501    /// This structure contains all the information needed to control the search,
502    /// beyond the energy function, the step function and the initial guess.
503    ///
504    /// - n_tries: The number of points to try for each step.
505    /// - iters: The number of iterations at each temperature.
506    /// - step_size: The maximum step size in the random walk.
507    /// - k, t_initial, mu_t, t_min: The parameters of the Boltzmann distribution and
508    ///   cooling schedule.
509    pub fn new(
510        n_tries: usize,
511        iters: usize,
512        step_size: f64,
513        k: f64,
514        t_initial: f64,
515        mu_t: f64,
516        t_min: f64,
517    ) -> SimAnnealingParams {
518        SimAnnealingParams {
519            n_tries,
520            iters_fixed_T: iters,
521            step_size,
522            k,
523            t_initial,
524            mu_t,
525            t_min,
526        }
527    }
528}