Introduction. Ride-hailing platforms have become a key component of urban service operations, and managing their efficiency falls within operations management: the real-time allocation of drivers to heterogeneous, uncertain demand. Despite this importance, systematic empirical evidence on these platforms in the Iranian market—and comparative evaluation of dispatch algorithms in the Persian literature—is scarce. This study first describes the Tehran ride-hailing market empirically using a large real dataset, then evaluates six dispatch algorithms (from simple rules to Hungarian optimization and reinforcement learning) in an agent-based simulation calibrated to the same data.

Methods. The study is applied in purpose and quantitative, simulation-based in method. Its empirical basis is 104770 real trips from an active Tehran ride-hailing platform over a 30-day window (20 March – 18 April 2024), anonymized before analysis; the descriptive analyses are performed directly on these data. Because of confidentiality, a representative synthetic dataset is released publicly instead of the raw data. An agent-based simulator (Mesa) was calibrated against five empirical patterns using the pattern-oriented modeling (POM) framework (completion-rate error 1.0%); three of five patterns fall inside the band and the weighted error is 16.5% against a 15% threshold, so POM is not fully met. Six dispatch algorithms were evaluated over 30 independent seeds and compared using the paired Wilcoxon test, effect sizes, bootstrap confidence intervals, the TOST equivalence test, and sensitivity analysis; passenger-experience equity was measured with the waiting-time Gini coefficient.

Results and discussion. The real anonymized data show an overall completion rate of 61.5%, a passenger-cancellation rate of 26.0%, and a no-driver rate of 12.5%; demand is highly non-uniform (peak-to-trough ratio about 33) and averages about 3,492 trips per day. A spatial analysis reflects a likely under-served sector (the south-east, 36.9% completion and a 2.8 km mean pickup distance, roughly three times the overall average), subject to field confirmation. Under heavy rain, the completion rate drops to 54.7% with no corresponding drop in driver acceptance. In the algorithmic evaluation, the salient finding is that the mathematically optimal batch-Hungarian algorithm degrades sharply under driver rejection — the realized offer-level rejection rate in the calibrated environment is 74.5% — (45.2% completion versus 62.5% for the greedy rule; p<0.001; d_z = 8.24; the greedy rule wins on all 30 seeds without exception); the sensitivity analysis shows this degradation is stable across the whole realistic acceptance range, with no level at which batch-Hungarian holds a significant advantage; a six-layer analysis of 6,570 runs, including seven-dimensional Latin-hypercube sampling and four dispatch-window lengths, confirms this, and adding a re-solving mechanism restores performance. Neither reinforcement-learning policy delivers a reliable operational improvement over the greedy rule: the double-Q network differs by −0.13 percentage points (one-sided p = 0.680) and TOST establishes its equivalence to greedy within ±0.015, while the value network shows a borderline +0.75 percentage points (one-sided p = 0.037, CI including zero); equivalence between the two architectures is not established. As a complementary equity signal, the double-Q network produces a lower waiting-time Gini (0.2984 versus 0.3105 for the greedy rule; p < 0.001) whereas the value network shows no significant difference; because of the waiting-time calibration gap this result is exploratory and relies only on the stability of the relative ranking. Collectively, the findings indicate that in dispatch-system design, robustness to acceptance uncertainty takes precedence over pure mathematical optimality.

Conclusions. From an operations-management standpoint, the mathematical optimality of a dispatch rule without accounting for acceptance uncertainty can be misleading; robustness to driver rejection is the decisive performance factor and should be the primary selection criterion. The main limitation is the single-city, simulation-based evaluation, calling for future field validation and multi-city extension.