BHAMSLE: A Breakpoint Heuristic Algorithm for Maximum Simulated Likelihood Estimation of Advanced Discrete Choice Models

Abstract

This paper introduces BHAMSLE, a Breakpoint Heuristic Algorithm for Maximum Simulated Likelihood Estimation (MSLE), adapted from the Breakpoint Heuristic Algorithm (BHA) for choicebased pricing, bridging the gap between choice-based optimization and choice model estimation. Similarly to the BHA, BHAMSLE leverages indifference points—or breakpoints—in individual decision-making to systematically explore local optima. Benchmark comparisons with PandasBiogeme, the current state-of-the-art software for DCM estimation, demonstrate that BHAMSLE, both as a standalone estimation procedure and as an approach for obtaining high-quality starting points, substantially improves log-likelihood on different latent class logit as well as latent class mixed logit models across 100 random samples, with gains of up to 10% for observed choices and up to 16% for synthetic choices. Notably, small numbers of draws are often enough to observe significant gains, with larger samples further amplifying these improvements.