1

Simulated Annealing
 Introduction 2

 The simulated annealing method is based on the simulation of thermal annealing

of critically heated solids.
 When a solid (metal) is brought into a molten state by heating it to a high
temperature, the atoms in the molten metal move freely with respect to each other.
 However, the movements of atoms get restricted as the temperature is reduced.
 As the temperature reduces, the atoms tend to get ordered and finally form crystals
having the minimum possible internal energy.
 Introduction 3

 The process of formation of crystals essentially depends on the cooling rate.

 When the temperature of the molten metal is reduced at a very fast rate, it may not
be able to achieve the crystalline state; instead, it may attain a polycrystalline state
having a higher energy state compared to that of the crystalline state.
 In engineering applications, rapid cooling may introduce defects inside the
material.
 Thus the temperature of the heated solid (molten metal) needs to be reduced at a
slow and controlled rate to ensure proper solidification with a highly ordered
crystalline state that corresponds to the lowest energy state (internal energy).
 This process of cooling at a slow rate is known as annealing.
 Procedure 4

 The simulated annealing method simulates the process of slow cooling of molten
metal to achieve the minimum function value in a minimization problem.
 The cooling phenomenon of the molten metal is simulated by introducing a
temperature-like parameter and controlling it using the concept of Boltzmann’s
probability distribution.
 The Boltzmann’s probability distribution implies that the energy (E) of a system in
thermal equilibrium at temperature T is distributed probabilistically according to
the relation

 P(E) = e−E/kT
 where P(E) denotes the probability of achieving the energy level E, and k is called
the Boltzmann’s constant.
 5

 Equation shows that at high temperatures the system has nearly a uniform
probability of being at any energy state;
 However, at low temperatures, the system has a small probability of being at a
high-energy state. This indicates that when the search process is assumed to follow
Boltzmann’s probability distribution, the convergence of the simulated annealing
algorithm can be controlled by controlling the temperature T .
 The method of implementing the Boltzmann’s probability distribution in simulated
thermodynamic systems, suggested by Metropolis et al. [13.37], can also be used
in the context of minimization of functions.
 In the case of function minimization, let the current design point (state) be Xi ,
with the corresponding value of the objective function given by fi = f (Xi ). Similar
to the energy state of a thermodynamic system, the energy Ei at state Xi is given by
6
 7

 Ei = fi = f (Xi )
 Then, according to the Metropolis criterion, the probability of the next design point
(state) Xi+1 depends on the difference in the energy state or function values at the two
design points (states) given by
 ∆E = Ei+1 − Ei = ∆f = fi+1 − fi ≡ f (Xi+1) − f (Xi )
 The new state or design point Xi+1 can be found using the Boltzmann’s probability
distribution:
 P[Ei+1] = min {1, e−∆E/kT }
 The Boltzmann’s constant serves as a scaling factor in simulated annealing and, as
such, can be chosen as 1 for simplicity. Note that if ∆E ≤ 0, Eq. (13.17) gives P[Ei+1] =
1and hence the point Xi+1 is always accepted. This is a logical choice in the context of
minimization of a function because the function value at Xi+1, fi+1, is better (smaller)
than at Xi , fi , and hence the design vector Xi+1 must be accepted. On the other hand,
when ∆E >0, the function value fi+1 at Xi+1 is worse (larger) than the one at Xi .
 8

 According to most conventional optimization procedures, the point Xi+1 cannot be

accepted as the next point in the iterative process. However, the probability of
accepting the point Xi+1, in spite of its being worse than Xi in terms of the objective
function value, is finite (although it may be small) according to the Metropolis
criterion. Note that the probability of accepting the point Xi+1
 P[Ei+1] ={ e−∆E/kT }(13.18)
 is not same in all situations. As can be seen from Eq. (13.18), this probability depends
on the values of ∆E and T. If the temperature T is large, the probability will be high for
design points Xi+1 with larger function values (with larger values of ∆E = ∆f). Thus, at
high temperatures, even worse design points Xi+1 are likely to be accepted because of
larger probabilities. However, if the temperature T is small, the probability of
accepting worse design points Xi+1 (with larger values of ∆E = ∆f) will be small. Thus,
as the temperature values get smaller (that is, as the process gets closer to the optimum
solution), the design points Xi+1 with larger function values compared to the one at Xi
are less likely to be accepted.
 9

 Start with an initial design vector X1 (iteration number i = 1) and a high value of temperature T .
 Generate a new design point randomly in the vicinity of the current design point and find the
difference in function values:
 ∆E = ∆f = fi+1 − fi ≡ f (Xi+1) − f (Xi ) (13.19)
 If fi+1 is smaller than fi (with a negative value of ∆f ), accept the point Xi+1 as the next design point.
Otherwise, when ∆f is positive, accept the point Xi+1 as the next design point only with a probability
e−∆E/kT . This means that if the value of a randomly generated number is larger than e−∆E/kT , accept the
point Xi+1; otherwise, reject the point Xi+1.
 If the point Xi+1 is rejected, then the process of generating a new design point Xi+1 randomly in the
vicinity of the current design point, evaluating the corresponding objective function value fi+1, and
deciding to accept Xi+1 as the new design point, based on the use of the Metropolis criterion, Eq.
(13.18), is continued.
 To simulate the attainment of thermal equilibrium at every temperature, a predetermined number (n) of
new points Xi+1 are tested at any specific value of the temperature T.
 10

 Once the number of new design points Xi+1 tested at any temperature T exceeds the value of n, the
temperature T is reduced by a prespecified fractional value c (0 < c < 1) and the whole process is
repeated.
 The procedure is assumed to have converged when the current value of temperature T is sufficiently
small or when changes in the function values (∆f ) are observed to be sufficiently small.
 The choices of the initial temperature T , the number of iterations n before reducing the temperature,
and the temperature reduction factor c play important roles in the successful convergence of the SA
algorithm.
 For example, if the initial temperature T is too large, it requires a larger number of temperature
reductions for convergence. On the other hand, if the initial temperature is chosen to be too small, the
search process may be incomplete in the sense that it might fail to thoroughly investigate the design
space in locating the global minimum before convergence.
 The temperature reduction factor c has a similar effect. Too large a value of c (such as 0.8 or 0.9)
requires too much computational effort for convergence.
 On the other hand, too small a value of c (such as 0.1 or 0.2) may result in a faster reduction in
temperature that might not permit a thorough exploration of the design space for locating the global
minimum solution.
 11

 Similarly, a large value of the number of iterations n will help in achieving quasi-
equilibrium state at each temperature but will result in a larger computational effort.
 A smaller value of n, on the other hand, might result either in a premature convergence
or convergence to a local minimum (due to inadequate exploration of the design space
for the global minimum).
 Unfortunately, no unique set of values are available for T , n, and c that will work well
for every problem. However, certain guidelines can be given for selecting these values.
 The initial temperature T can be chosen as the average value of the objective function
computed at a number of randomly selected points in the design space.
 The number of iterations n can be chosen between 50 and 100 based on the computing
resources and the desired accuracy of solution. The temperature reduction factor c can
be chosen between 0.4 and 0.6 for a reasonable temperature reduction strategy (also
termed the cooling schedule).
Features of Simulated Annealing 12

 Some of the features of simulated annealing are as follows:

 1. The quality of the final solution is not affected by the initial guesses, except that the
computational effort may increase with worse starting designs.
 2. Because of the discrete nature of the function and constraint evaluations, the
convergence or transition characteristics are not affected by the continuity or
differentiability of the functions.
 3. The convergence is also not influenced by the convexity status of the feasible space.
 4. The design variables need not be positive.
 5. The method can be used to solve mixed-integer, discrete, or continuous problems.
 6. For problems involving behavior constraints (in addition to lower and upper bounds
on the design variables), an equivalent unconstrained function is to be formulated as in
the case of genetic algorithms
13
14
15