Exploring the potential of computational graph-based gradients for choice modelling

Choice modelling is widely used to analyse travel behaviour, but increasing model complexity leads to estimation challenges including increased model run times and multiple local optima. Computational Graph (CG) offers quick and accurate approximation of the likelihood function’s gradient, thereby addressing a key limitation of traditional gradient calculation methods. This study contrasts the performance of CG-based, analytical, and numerical gradient calculation methods for latent class and mixed logit models. Our findings highlight that CG achieves precise gradient estimates whilst significantly reducing estimation time. Analytical and CG-based gradient methods are less likely to result in bad local optima compared to numerical derivatives when testing across a wide range of starting values. Although local optima still occur, CG’s faster estimation allows feasible testing over a range of starting values. As such, it represents a valuable tool, given the significant implications of poor local optima in terms of key model outputs.