Using a Simulated Annealing Algorithm to Generate Efficient Stated Preference Designs
P Davidson, C Teye-Ali, R Culley, Peter Davidson Consultancy, UK
In this paper we explore the scope for using Simulated Annealing a new method of generating efficient designs. This algorithm is compared with existing methods in terms of execution time, the quality of the designs produced
Stated preference has limited scope for including different variables at different levels so it is important to make the best use of the few variables which can be included in the design and ensuring that their levels operate at maximum efficiency. The traditional approach was to use an orthogonal fractional factorial design but this places severe limitations on the number of variables and their levels. A better way can be to forego some of the orthogonality in favour of more variables and more levels. This can use a measure of efficiency which is optimised to provide the best design. These measures include the "A Error", "D Error" and "G Error" and the task is to find the "Efficient" design with the minimum error.
There are various different heuristic algorithms for generating the efficient design and this paper explores these different methods. They generally use heuristic algorithms including: Local search, Fedorev exchange or modified Fedorev exchange. In our search for better heuristic algorithms, we have investigated the scope for using Simulated Annealing which we understand has not hitherto been investigated. In this paper we explore the scope for using Simulated Annealing in solving this problem. This is compared in this paper to the above existing methods. These algorithms can take a long time to converge so this paper will outline their speed of convergence, execution time, the quality of the design produced together with the various measures of error when compared with Simulated Annealing.
This comparative exercise fed into the design of a stated and revealed preference exercise in Calabar, Nigeria to measure the logit model coefficients for a lot of key variables against a background where none of these had been measured before including: in-vehicle time, fare, walk time, wait time, interchange penalty, mode perception, cost for the various different public transport modes including large bus, minibus, shared taxi as well as for private vehicles. The paper concludes with a comparison of how the chosen designs worked out in practice when it comes to providing coefficient estimates.
Association for European Transport