Sequential Experiment Designs for Screening and Tuning Parameters of Stochastic Heuristics

This paper describes a sequential experimentation approach for efficiently screening and tuning the parameters of a stochastic heuristic. Stochastic heuristics such as ant colony algorithms often use a large number of tuning parameters. Testing all combinations of these factors is prohibitive and inefficient. The sequential procedure recommended by this paper uses resolution IV fractional factorial designs with fold-over and centre points as an efficient way to screen the most important tuning parameters. The effects of the most important parameters are then modelled using a central composite design and optimised with standard numerical methods. All designs, their analyses and interpretation are illustrated using the Ant Colony System algorithm. The use of standard designs and methods has the benefit that the presented procedure can easily be followed with commercial software rather that relying on custom methodologies and tools that have only been developed in an academic context. Such a procedure has not been applied to ant colony algorithms before.