Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Evaluate sparse retrieval against feasible all-pairs search at scale

Determine the isolated live-electric-vehicle quality effect of the sparse request–vehicle retriever by comparing it with a feasible all-pairs candidate-generation control for dynamic electric ride-pooling instances containing 2,000–8,000 requests.

Background

The L2R-EV framework reduces the quadratic request–vehicle candidate space by retaining 16 target vehicles per request: ten selected through learned similarity, four nearby vehicles, and two random exploratory targets. Every retained candidate is subsequently checked for exact route, passenger, battery, and immediate-saving feasibility. Large-scale experiments demonstrate end-to-end execution, but they do not isolate whether sparse retrieval itself affects live-EV solution quality because the experiments lack a feasible comparison against complete all-pairs candidate generation.

Resolving this problem would require constructing a computationally feasible all-pairs control at the reported large scales, or otherwise developing an equivalent evaluation design that separates the effect of sparse candidate retrieval from the effects of the learned repair scorer, exact-test budget, and batching rules.

References

The sparse retriever is required to avoid constructing the large all-pairs matrix, but we do not have a feasible all-pairs control at $n=2000$--8000 and therefore do not claim an isolated live-EV quality effect for retrieval.

— L2R-EV: Learning What to Repair in Electric Ride-Pooling with Finite Charger Queues  (2609.40225 - Pham et al., 30 Sep 2026) in Appendix, Section Component ablations