Simulated Annealing. Simulated Annealing is inspired by the process of metallurgical annealing: when a metal is heated, the atoms are free to change their position randomly in the high-energy environment. By controlling how quickly the metal cools, it is possible to encourage the atoms in the metal to form in stronger, more beneficial arrangements, thus strengthening the metal. Simulated annealing works in a similar way: • An initial solution is generated and a high 'global temperature' is set. • The solution is randomly mutated to another solution by making a single change. • Acceptance of the solution depends on the temperature; if it is high, then highly random changes are possible and acceptable; if it is low, then only minor changes are possible. • The process repeats but the global temperature is steadily reduced. The result of the process is that large areas of the design space are explored early on via highly random changes, and later on only minor improvements are made. Simulated annealing follows a greedy method - it seeks to achieve a global optimum by making decisions that are locally optimum at each stage [15]. Simulated xxxxxxxxx's primary advantage is that it is very fast, but as a result it is not always as thorough as other algorithms.
Simulated Annealing. Simulated Annealing is related to Hill Climbing, but it attempts to avoid being trapped by local maxima by probabilistically being able to jump to a neighbour with a lower value if one is found. The probability of this jump to a seemingly worse design is initially quite high (>0.5) but as the number of simulations performed increases, the probability gradually decreases. The optimum design is either the one being examined at the end of the experiment when all simulations have been run or possibly an earlier one if the records of all designs are kept and there were some jumps to lower values near the end of the experiment. The DSE driver needs to know the allowed values for the design parameters, the rate at which the probability of jumping decreases along with the total number of simulations to perform, and access to the simulation results.
