H-TD²: Hybrid Temporal Difference Learning for Adaptive Urban Taxi Dispatch

We present H-TD 2 : Hybrid Temporal Difference Learning for Taxi Dispatch, a model-free, adaptive decision-making algorithm to coordinate large fleet of automated taxis in dynamic urban environment minimize expected customer waiting times. Our scalable exploits the natural transportation network company topology by switching between two behaviors: distributed temporal-difference learning computed locally at each taxi and infrequent centralized Bellman updates dispatch center. derive regret bound design trigger condition behaviors explicitly control trade-off computational complexity individual policy’s bounded sub-optimality; this advances state art enabling operation with bounded-suboptimality. Additionally, unlike recent reinforcement methods, policy estimation is robust out-of-training domain events. This result enabled two-step modelling approach: learned on an agent-agnostic, cell-based Markov Decision Process are coordinated using game-theoretic task assignment. validate our against receding horizon baseline Gridworld simulated dataset, where proposed solution decreases average time 50% over wide range parameters. also Chicago city real requests from public dataset 26% irregular distributions during 2016 Major League Baseball World Series game.