Skip to main content
Back to timeline
arXivSource publication:

Ride-hailing repositioning trains its demand predictor on decisions: 7.7% to 9.1% more business benefit than mean-squared-error training in New York City data

Synopsis

A ride-hailing platform forecasts demand, executes repositioning, and carries the resulting fleet state into the next stage. Demand predictors were previously trained on mean squared error, disconnected from the scheduling decision; multistage contextual optimization instead trains the predictor on the operating cost of the whole trajectory and uses the state left by earlier decisions as prediction context. On New York City ride-hailing data, the same predictor trained on decision cost yields 7.7% to 9.1% more business benefit than mean-squared-error training and outperforms all compared baselines from five paradigms.

Source-provided article image: Multistage Contextual Optimization for Vehicle Repositioning
arXiv

Interpretation

The demand predictor for idle-vehicle repositioning is no longer trained only for numerical forecast accuracy; it is trained on the actual operating cost of the whole multistage scheduling trajectory. Such predictors were previously trained on forecast error such as mean squared error, while the cost of a forecast error depends on where idle vehicles are, and their locations depend on earlier repositioning and trips. On New York City ride-hailing data, the same predictor trained on decision cost yields 7.7% to 9.1% more business benefit than mean-squared-error training, is higher on all 20 test days, and outperforms all compared baselines from five paradigms.

The training signal must pass through several coupled scheduling stages: the current forecast changes the objective, constraint matrix, and right-hand side of the linear program, while the fleet state left by the previous stage in turn serves as prediction context. Existing multistage contextual optimization studies propagate gradients across stages but do not capture how the inherited state changes the forecast as context, and one of them restricts uncertainty to the objective. The exact gradient has three parts: parameters change forecasts and hence decisions, earlier decisions affect the current decision through the inherited state, and this state also changes forecasts as context; an equivalent adjoint uses reverse vector–Jacobian products to avoid full solution and state-parameter Jacobians.

Forward propagation solves many stage linear programs one after another, the main training bottleneck; parallel block evaluation solves blocks of consecutive stages in parallel and repeatedly feeds each block output to the next block. Existing acceleration methods reduce the cost of differentiation or the number of solver calls but do not shorten the sequential forward evaluation of one sample. With eight workers, parallel block evaluation shortens training by a factor of 1.86 without loss of decision performance; 30 epochs take 19.9 hours versus 37.1 hours for the serial version.

Perspective

The evaluation covers New York City ride-hailing data with the city partitioned into eight regions and 64 regional OD pairs, 180 decision stages per day from 06:00 to 21:00, and results come from an agent-based simulation rather than a live platform deployment.

The gradient is exact only in a neighborhood where every linear program keeps its optimal basis and every rule stays on its branch; near a basis change the gradient can be orders of magnitude larger, and such updates are skipped, 35 of 930 visits for the serial version and 152 for the parallel version.