Exact Estimation of Activity-Based Dynamic Discrete Choice Models via Reachability-Based Sparse Computation

Abstract

Dynamic Discrete Choice Models (DDCMs) provide a theoretically rigorous, welfare-consistent foundation for activity-based travel demand modeling, but the curse of dimensionality has prevented exact estimation at a practical scale. This study develops a GPU-accelerated computational framework that resolves this barrier through three contributions: reachability-based state space pruning that eliminates the vast majority of physically unreachable states, an activity-sequence partitioned universal graph with constraint-masked batched backward induction that handles population heterogeneity without per-person graph construction, and a time-varying marginal utility specification in which activity timing and sequencing emerge from preference gradients alone. The framework is demonstrated using travel diary data from Higashi-Hiroshima, Japan, with exact NFXP estimation on a single GPU.