A dynamic differential evolution algorithm for mixed logit discrete choice model estimation

The mixed logit (ML) discrete choice model is highly flexible and capable of modeling complex choice behaviors. A popular method for estimation of an ML model is through maximization of a simulated likelihood function, which, however, often contains multiple local optima in a high-dimensional solution space. This paper reports the development of a dynamic differential evolution (DE) algorithm for the estimation of a general ML model with correlated tastes and repeated choices. Compared with the gradient based algorithms that are commonly adopted in literature, the proposed DE algorithm is less sensitive to the properties of the distributions assumed and the conditions of initialization, and it is more robust in converging to near optimal solutions.