Automating Parameter Choice For Simulated Annealing

Most metaheuristic techniques, including Simulated Annealing, require the specification of parameters before they can be used. Setting these parameters is not straightforward, and there is no accepted way to determine good values for the parameters other than trial and error, which is uncertain and can be very time-consuming. Experimental results are presented to help determine how parameters for Simulated Annealing should be set automatically, given only the time available for a solution. These results, for two randomly-generated Euclidean Travelling Salesman Problems, demonstrate, among other features, that the prevailing orthodoxy concerning starting temperatures is significantly suboptimal. The paper ends with a suggested set of steps to be undertaken to determine the parameters for any given instance