Adapting service behavior of taxi drivers for optimal operating decisions via distributional reinforcement learning

Taxi drivers’ area choices shape subsequent locations, order availability, and cumulative earnings, yet previous intelligent transport studies prioritize passenger-demand matching or expected returns, overlooking sequential dependence and heterogeneous risk preferences. This study formulates taxi dispatch as a risk-sensitive, multi-step decision problem to support taxi drivers’ adaptive service behavior and proposes a weighted implicit quantile network (wIQN) to optimize long-horizon operating revenue. wIQN extends the standard implicit quantile network with target-quantile-aware temporal-difference weighting and a correction for extreme negative weights, thereby preserving informative variation in the return distribution. It further separates preference-neutral distribution learning through uniform quantile sampling from preference-specific policy selection through quantile distortion. We evaluate wIQN in a week-long offline simulation built from 2019 New York City taxi trips, comparing it with greedy, deep Q-network (DQN), and implicit quantile network (IQN) policies under risk-averse, risk-neutral, and risk-seeking preferences. The results show that wIQN achieves the strongest weekly revenue performance, learns broader and more differentiated return distributions than IQN, and produces clearer responses to heterogeneous risk preferences. These findings indicate that preserving return-distribution detail during training and postponing preference distortion until decision time can improve long-horizon decisions while supporting personalized dispatch recommendations. The study contributes an operator-centered, risk-sensitive distributional reinforcement-learning framework for the model-driven adaptation of service behavior in dynamic urban service systems.