Simulated annealing copies a phenomenon in nature--the annealing of solids--to optimize a complex system. Annealing refers to heating a solid and then cooling it slowly. Atoms then assume a nearly globally minimum energy state. In 1953 Metropolis created an algorithm to simulate the annealing process. The algorithm simulates a small random displacement of an atom that results in a change in energy. If the change in energy is negative, the energy state of the new configuration is lower and the new configuration is accepted. If the change in energy is positive, the new configuration has a higher energy state; however, it may still be accepted according to the Boltzmann probability factor:

where k_{b} is the Boltzmann constant and T is the current temperature. By examining this equation we should note two things: the probability is proportional to temperature--as the solid cools, the probability gets smaller; and inversely proportional to --as the change in energy is larger the probability of accepting the change gets smaller.

When applied to engineering design, an analogy is made between energy and the objective function. The design is started at a high “temperature”, where it has a high objective (we assume we are minimizing). Random perturbations are then made to the design. If the objective is lower, the new design is made the current design; if it is higher, it may still be accepted according the probability given by the Boltzmann factor. The Boltzmann probability is compared to a random number drawn from a uniform distribution between 0 and 1; if the random number is smaller than the Boltzmann probability, the configuration is accepted. This allows the algorithm to escape local minima.

As the temperature is gradually lowered, the probability that a worse design is accepted becomes smaller. Typically at high temperatures the gross structure of the design emerges which is then refined at lower temperatures.

Although it can be used for continuous problems, simulated annealing is especially effective when applied to combinatorial or discrete problems. Although the algorithm is not guaranteed to find the best optimum, it will often find near optimum designs with many fewer design evaluations than other algorithms. (It can still be computationally expensive, however.) It is also an easy algorithm to implement.

You can download anneal.m and anneal.py files to retrieve example simulated annealing files in MATLAB and Python, respectively.

Download Simulated Annealing Example Files

comments powered by Disqus

Retrieved from http://apmonitor.com/me575/index.php/Main/SimulatedAnnealing

Page last modified on May 31, 2018, at 09:03 PM