optim_sa

Arguments

x0

the initial solution

f

the objective function to be optimized (the prototype if f(x))

ItExt

the number of temperature decrease

ItInt

the number of iterations during one temperature stage

T0

the initial temperature (see compute_initial_temp to compute easily this temperature)

Log

if %T, some information will be displayed during the run of the simulated annealing

temp_law

the temperature decrease law (see temp_law_default for an example of such a function)

param_temp_law

a structure (of any kind - it depends on the temperature law used) which is transmitted as a parameter to temp_law

neigh_func

a function which computes a neighbor of a given point (see neigh_func_default for an example of such a function)

param_neigh_func

a structure (of any kind like vector, list, it depends on the neighborhood function used) which is transmitted as a parameter to neigh_func

x_best

the best solution found so far

f_best

the objective function value corresponding to x_best

mean_list

the mean of the objective function value for each temperature stage. A vector of float (optional)

var_list

the variance of the objective function values for each temperature stage. A vector of float (optional)

f_history

the computed objective function values for each iteration. Each input of the list corresponds to a temperature stage. Each input of the list is a vector of float which gathers all the objective function values computed during the corresponding temperature stage - (optional)

temp_list

the list of temperature computed for each temperature stage. A vector of float (optional)

x_history

the parameter values computed for each iteration. Each input of the list corresponds to a temperature stage. Each input of the list is a vector of input variables which corresponds to all the variables computed during the corresponding temperature stage - (optional - can slow down a lot the execution of optim_sa)

iter

a double, the actual number of external iterations in the algorithm (optional).

Description

A Simulated Annealing optimization method.

Simulated annealing (SA) is a generic probabilistic meta-algorithm for the global optimization problem, namely locating a good approximation to the global optimum of a given function in a large search space. It is often used when the search space is discrete (e.g., all tours that visit a given set of cities).

The current solver can find the solution of an optimization problem without constraints or with bound constraints. The bound constraints can be customized with the neighbour function. This algorithm does not use the derivatives of the objective function.

The solver is made of Scilab macros, which enables a high-level programming model for this optimization solver. The SA macros are based on the parameters Scilab module for the management of the (many) optional parameters.

To use the SA algorithm, one should perform the following steps :

• configure the parameters with calls to init_param and add_param especially the neighbor function, the acceptance function, the temperature law,
• compute an initial temperature with a call to compute_initial_temp,
• find an optimum by using the optim_sa solver.

The algorithm is based on an iterative update of two points :

• the current point is updated by taking into account the neighbour and the acceptance functions,
• the best point is the point which achieved the minimum of the objective function over the iterations.

The acceptance of the new point depends on the output values produced by the rand function. This implies that two consecutive calls to the optim_sa will not produce the same result. In order to always get exactly the same results, please initialize the random number generator with a valid seed.

See the Demonstrations, in the "Optimization" section and "Simulated Annealing" subsection for more examples.

The objective function

The objective function is expected to have the following header.

function y=f(x)

In the case where the objective function needs additionnal parameters, the objective function can be defined as a list, where the first argument is the cost function, and the second argument is the additionnal parameter. See below for an example.

A simple example

In the following example, we search the minimum of the Rastriging function. This function has many local minimas, but only one single global minimum located at x = (0,0), where the function value is f(x) = -2. We use the simulated annealing algorithm with default settings and the default neighbour function neigh_func_default.

function y=rastrigin(x)
  y = x(1)^2+x(2)^2-cos(12*x(1))-cos(18*x(2));
endfunction
    
x0          = [2 2];
Proba_start = 0.7;
It_Pre      = 100;
It_extern   = 100;
It_intern   = 1000;
x_test = neigh_func_default(x0);

T0 = compute_initial_temp(x0, rastrigin, Proba_start, It_Pre);

Log = %T;
[x_opt, f_opt, sa_mean_list, sa_var_list] = optim_sa(x0, rastrigin, It_extern, It_intern, T0, Log);

printf("optimal solution:\n"); disp(x_opt);
printf("value of the objective function = %f\n", f_opt);

t = 1:length(sa_mean_list);
plot(t,sa_mean_list,"r",t,sa_var_list,"g");

Configuring a neighbour function

In the following example, we customize the neighbourhood function. In order to pass this function to the optim_sa function, we setup a parameter where the "neigh_func" key is associated with our particular neighbour function. The neighbour function can be customized at will, provided that the header of the function is the same. The particular implementation shown below is the same, in spirit, as the neigh_func_default function.

function f=quad(x)
  p = [4 3];
  f = (x(1) - p(1))^2 + (x(2) - p(2))^2
endfunction

// We produce a neighbor by adding some noise to each component of a given vector
function x_neigh=myneigh_func(x_current, T, param)
  nxrow = size(x_current,"r")
  nxcol = size(x_current,"c")
  sa_min_delta = -0.1*ones(nxrow,nxcol);
  sa_max_delta = 0.1*ones(nxrow,nxcol);
  x_neigh = x_current + (sa_max_delta - sa_min_delta).*rand(nxrow,nxcol) + sa_min_delta;
endfunction

x0          = [2 2];
Proba_start = 0.7;
It_Pre      = 100;
It_extern   = 50;
It_intern   = 100;

saparams = init_param();
saparams = add_param(saparams,"neigh_func", myneigh_func);
// or: saparams = add_param(saparams,"neigh_func", neigh_func_default);
// or: saparams = add_param(saparams,"neigh_func", neigh_func_csa);
// or: saparams = add_param(saparams,"neigh_func", neigh_func_fsa);
// or: saparams = add_param(saparams,"neigh_func", neigh_func_vfsa);

T0 = compute_initial_temp(x0, quad, Proba_start, It_Pre, saparams);
Log = %f;
// This should produce x_opt = [4 3]
[x_opt, f_opt] = optim_sa(x0, quad, It_extern, It_intern, T0, Log,saparams);

Passing extra parameters

In the following example, we use an objective function which requires an extra parameter p. This parameter is the second input argument of the quadp function. In order to pass this parameter to the objective function, we define the objective function as list(quadp,p). In this case, the solver makes so that the calling sequence includes a second argument.

function f=quadp(x, p)
f = (x(1) - p(1))^2 + (x(2) - p(2))^2
endfunction

x0 = [-1 -1];
p = [4 3];
T0 = compute_initial_temp(x0, list(quadp,p) , Proba_start, It_Pre);
[x_opt, f_opt] = optim_sa(x0, list(quadp,p) , 10, 1000, T0, %f);

Configuring an output function

In the following example, we define an output function, which also provide a stopping rule. We define the function outfun which takes as input arguments the data of the algorithm at the current iteration and returns the boolean stop. This function prints a message into the console to inform the user about the current state of the algorithm. It also computes the boolean stop, depending on the value of the function. The stop variable becomes true when the function value is near zero. In order to let optim_sa know about our output function, we configure the "output_func" key to our outfun function and call the solver. Notice that the number of external iterations is %inf, so that the external loop never stops. This allows to check that the output function really allows to stop the algorithm.

function f=quad(x)
  p = [4 3];
  f = (x(1) - p(1))^2 + (x(2) - p(2))^2
endfunction

function stop=outfunc(itExt, x_best, f_best, T, saparams)
  [mythreshold,err] = get_param(saparams,"mythreshold",0);
  mprintf ( "Iter = #%d, \t x_best=[%f %f], \t f_best = %e, \t T = %e\n", itExt , x_best(1),x_best(2) , f_best , T )
  stop = ( abs(f_best) < mythreshold )
endfunction

x0 = [-1 -1];
saparams = init_param();
saparams = add_param(saparams,"output_func", outfunc );
saparams = add_param(saparams,"mythreshold", 1.e-2 );

rand("seed",0);

T0 = compute_initial_temp(x0, quad , 0.7, 100, saparams);
[x_best, f_best, mean_list, var_list, temp_list, f_history, x_history , It ] = optim_sa(x0, quad , %inf, 100, T0, %f, saparams);

The previous script produces the following output. Notice that the actual output of the algorithm depends on the state of the random number generator rand: if we had not initialize the seed of the uniform random number generator, we would have produced a different result.

Iter = #1,      x_best=[-1.000000 -1.000000],      f_best = 4.100000e+001,      T = 1.453537e+000
Iter = #2,      x_best=[-0.408041 -0.318262],      f_best = 3.044169e+001,      T = 1.308183e+000
Iter = #3,      x_best=[-0.231406 -0.481078],      f_best = 3.002270e+001,      T = 1.177365e+000
Iter = #4,      x_best=[0.661827 0.083743],      f_best = 1.964796e+001,      T = 1.059628e+000
Iter = #5,      x_best=[0.931415 0.820681],      f_best = 1.416565e+001,      T = 9.536654e-001
Iter = #6,      x_best=[1.849796 1.222178],      f_best = 7.784028e+000,      T = 8.582988e-001
Iter = #7,      x_best=[2.539775 1.414591],      f_best = 4.645780e+000,      T = 7.724690e-001
Iter = #8,      x_best=[3.206047 2.394497],      f_best = 9.969957e-001,      T = 6.952221e-001
Iter = #9,      x_best=[3.164998 2.633170],      f_best = 8.317924e-001,      T = 6.256999e-001
Iter = #10,      x_best=[3.164998 2.633170],      f_best = 8.317924e-001,      T = 5.631299e-001
Iter = #11,      x_best=[3.164998 2.633170],      f_best = 8.317924e-001,      T = 5.068169e-001
Iter = #12,      x_best=[3.961464 2.903763],      f_best = 1.074654e-002,      T = 4.561352e-001
Iter = #13,      x_best=[3.961464 2.903763],      f_best = 1.074654e-002,      T = 4.105217e-001
Iter = #14,      x_best=[3.961464 2.903763],      f_best = 1.074654e-002,      T = 3.694695e-001
Iter = #15,      x_best=[3.931929 3.003181],      f_best = 4.643767e-003,      T = 3.325226e-001

See Also

• compute_initial_temp
• neigh_func_default
• temp_law_default

Authors

2008, Yann COLLETTE (ycollet@freesurf.fr)

Michael Baudin - Digiteo - 2010

Bibliography

"Simulated annealing : theory and applications", P.J.M. Laarhoven and E.H.L. Aarts, Mathematics and its applications, Dordrecht : D. Reidel, 1988

"Theoretical and computational aspects of simulated annealing", P.J.M. van Laarhoven, Amsterdam, Netherlands : Centrum voor Wiskunde en Informatica, 1988

"Genetic algorithms and simulated annealing", Lawrence Davis, London : Pitman Los Altos, Calif. Morgan Kaufmann Publishers, 1987