PaperlayerView exact revision on arXiv

Source-linked AI summary

L2R-EV: Learning What to Repair in Electric Ride-Pooling with Finite Charger Queues

Mai Pham, Vikrant S. Vaze, Peter Chin

arXiv:2609.40225v1cs.GT

TL;DR

Electric ride-pooling must coordinate passenger service, batteries, and finite chargers, while a locally useful route change can reduce later service. L2R-EV combines a discrete-event simulator with learned passenger reassignment and a policy that directs exact testing. It improves held-out service, reduces exact testing, and scales to large episodes; queue-aware charging also cuts waiting.

  • Problem

    Dynamic pooled EV service must preserve route feasibility amid battery constraints and finite shared chargers, while accounting for whether immediate repairs harm later service.

    Source: §1 · pp. 1–2

  • Method

    L2R-EV combines discrete-event simulation with learned passenger reassignment, supervised future-effect scoring, and two-stage PPO to allocate exact tests.

    Source: §1 · p. 2, §3 · pp. 2–3

  • Results

    Supervised repair improves held-out EV service over no repair and rolling reoptimization; PPO reduces exact evaluations from 1,530 to 70 per episode, and sparse repair runs through 8,000 requests.

    Source: §5 · p. 4, §4 · p. 4

  • Takeaways & Limitations

    Current-workload charging cuts queue time by 41.0% versus nearest charging with greedy dispatch; at 1.5 kW, operational electricity per served request falls 1.73%.

    Source: §4 · p. 4

Abstract

from arXiv · show

Electric ride-pooling must jointly manage passenger service, routes, batteries, and finite chargers; a locally useful relocation can reduce later service. We introduce a discrete-event ride-pooling simulator with ordered passenger stops, pickup and ride-time constraints, battery reserves, charger travel, and finite first-come, first-served (FCFS) charging queues. On top of a common dispatcher, we add a learned repair layer that moves uncollected passengers between vehicles, uses a supervised score to estimate later effects, and uses a two-stage proximal-policy-optimization (PPO) policy to focus computation on promising moves. Every executed move must satisfy route, battery, and time constraints and provide an immediate route-cost improvement. For larger instances, we limit the number of candidate moves and apply several non-conflicting repairs at once. On 12 held-out Manhattan episodes with 100 requests, 30 EVs, and one plug per station, supervised repair reduces operational cost by 6.57\% versus no repair (12/12 wins) and serves 74.92 rather than 72.75 requests. Two-stage PPO stays within 1.43\% of supervised repair with 95.4\% fewer exact trials, while our experiments scale to 8,000 requests and 2,400 EVs with cost reductions across all three larger development settings. Queue-aware charging reduces waiting by 41.0\%; under a declared 1.5-kW queue-idle load, operational electricity per served request falls 1.73\%.

1 Introduction

Dynamic electric ride-pooling must balance passenger service, route feasibility, and finite shared chargers; the paper introduces a simulator and learned repair approach for this setting.

1 Introduction

  • Pooling and electrification do not by themselves guarantee lower system-wide emissions, so evaluation must connect passenger service with distance, energy, and charging.Source: §1 · p. 1
  • Dynamic requests arrive amid ordered passenger commitments, making each update balance time limits, capacity, battery reserve, charger access, and future demand.Source: §1 · p. 1
  • The open intersection is dynamic, high-capacity pooled service with shared-plug queues, where learned guidance must preserve exact route feasibility.Source: §1 · p. 2
  • The work contributes an event-driven EV ride-pooling simulator, learned repair with exact checks, and evaluation through 8,000 requests.Source: §1 · p. 2

2 Dynamic EV simulator

The simulator represents evolving passenger, vehicle, and charger states, enforces service and battery constraints, and evaluates a weighted operational cost.

2 Dynamic EV simulator

  • At each time, the simulator observes released requests, ordered stops, vehicle state, state of charge, charger state, and station queues.Source: §2 · p. 2
  • Routes must satisfy pickup-wait, ride-time, passenger-capacity, state-of-charge, and reserve constraints; the composite cost uses fixed weights for five operational terms.The terms are rejection, vehicle time, energy, passenger wait, and excess ride time; lower cost is better, while charger queue time remains a reported mechanism metric.Source: §2 · p. 2
  • Stations have identical plugs and deterministic first-come, first-served queues; waiting vehicles receive no energy and charging completion is an event.A vehicle releases its plug when charging ends or passenger dispatch makes it leave; comparisons use the same request stream and initial fleet.Source: §2 · p. 2

3 L2R-EV: Learning What to Repair in Electric Ride-Pooling

L2R-EV proposes passenger reassignments, verifies selected moves exactly, and learns both which moves may help later service and which proposals merit testing.

3 L2R-EV: Learning What to Repair in Electric Ride-Pooling

  • A repair reassigns an accepted passenger before pickup; the simulator recomputes both routes and accepts only feasible moves that reduce their remaining cost.Rejected tests restore the original routes and event times exactly.Source: §3 · p. 2
  • A supervised future-score network learns from paired simulations with identical states and future requests, labeling the later cost difference between accepting a move and doing nothing.The proposal-ranking network limits exact tests to at most 96 per repair time; inputs exclude unreleased requests, and a validation-set threshold governs acceptance.Source: §3 · p. 3
  • PPO allocates exact tests by resolving four of 16 proposals and may reject a tested move using its known route saving.Training uses final episode cost to value later passenger and charging effects.Source: §3 · p. 3
  • At large scale, each request keeps 16 candidate vehicles—10 similarity-selected, four nearby, and two random—and vehicle-disjoint moves can execute in batches.Batch size is capped at the ceiling of fleet size divided by 150.Source: §3 · p. 3
  • Low-charge vehicles choose the nearest charger or the station with the least travel time plus current workload; queue-idle energy is modeled as 1.5 kW times queue hours.Source: §3 · p. 3

4 Results

Supervised repair improves held-out results, while PPO approaches its cost with fewer exact evaluations. The approach also retains gains with finite charging capacity and scales to larger fleets, while queue-aware charging reduces waiting and operational electricity per served request.

Solution quality and exact tests

  • PPO before and after testing gives cost within 1.43% of supervised repair while reducing exact evaluations from 1,530 to 70 per episode.At matched work, PPO improves over both random and proposal-score allocation.Source: §4 · pp. 3–4, Table 1 · p. 3
  • Supervised repair improves cost on all 12 seeds; against rolling reoptimization with the same workload-aware charging rule, it lowers aggregate cost by 25.06%, wins 11/12 episodes, and serves 11.0 more requests on average.The main experiment uses 12 paired held-out Manhattan episodes with 100 requests, 30 four-seat EVs, a 1.5-hour window, and one plug per station.Source: §4 · pp. 3–4, Table 1 · p. 3

Finite-capacity robustness

  • With one plug, supervised repair still wins on all 12 paired episodes while operating through FCFS queues.The same trained model reduces cost from one plug through unlimited charging; four and eight plugs produce identical routes and zero waiting.Source: §4 · p. 4, Table 2 · p. 4

Large-scale experiments

  • The large-scale configuration lowers cost versus no repair in 9 of 10 episodes and versus one move per epoch in all 10, while exact work stays near 1,700 tests as requests grow fourfold.The share of request–vehicle combinations proposed falls from 2.67% to 0.67%, and the share tested exactly falls from 0.130% to 0.008%.Source: §4 · p. 4, Table 3 · p. 4

Queue waiting and operational energy

  • Current-workload charging cuts queue time by 41.0% versus nearest charging with greedy dispatch and 36.0% with reoptimization.At 1.5 kW, operational electricity per served request falls 1.73% (0.642 versus 0.653 kWh; [0.54, 2.98]%, p = 0.0186).Source: §4 · p. 4

5 Conclusion and outlook

The conclusion emphasizes exact feasibility, improved service and cost, reduced PPO testing, and dynamic execution at large scale, while identifying future validation and modeling needs.

5 Conclusion and outlook

  • Exact simulation preserves feasibility, supervised repair improves held-out EV service over no repair and rolling reoptimization, PPO reduces exact testing, and sparse vehicle-disjoint repair completes dynamic episodes with 8,000 requests.Future work will confirm large-scale rules across more datasets and develop batch-aware future scores, charger policies valuing downstream passenger service, richer energy physics, and time-varying grid emissions.Source: §5 · p. 4

A Environment, data, and comparison methods

The simulator models event-driven passenger service, vehicle routes, batteries, and finite charger queues. Its learned repair pipeline evaluates feasible passenger reassignments, estimates later effects, and uses sparse retrieval and nonconflicting batches at scale.

A Environment, data, and comparison methods

  • The simulator advances to the next request, stop, charging-completion, or decision event, tracking released requests and each vehicle’s position, ordered stops, load, SOC, and charging or queue status.Unreleased requests are not visible to a policy; actions can assign or reject requests, reassign accepted passengers before pickup, reposition idle vehicles, or send low-SOC vehicles to charge.Source: A Environment, data, and comparison… · p. 6
  • Routes and repairs enforce a 300-second pickup deadline, ride time capped at twice direct travel time, passenger capacity, and a 5-kWh battery reserve; passengers already aboard cannot be moved.The modeled EV has a 62-kWh battery, 0.15 kWh/km traction consumption, and 72-kW charging.Source: A Environment, data, and comparison… · p. 6
  • Each station uses identical plugs and deterministic FCFS queues; waiting vehicles receive no energy, and a plug is released when charging ends or passenger dispatch makes a vehicle leave.Workload-aware charging minimizes travel time plus visible unfinished charging work, but does not predict future requests or charger arrivals.Source: A Environment, data, and comparison… · p. 6
  • Manhattan episodes sample weekday demand and travel data, while policies in each pair share demand, initial vehicle positions, and initial SOC but evolve independently after actions differ.Held-out n = 100 episodes have realized demand from 42 to 123; large-scale experiments use exactly the stated request count to isolate computational scale from demand variation. The perfect-information reference is capacity-relaxed because it does not schedule plug competition.Source: A Environment, data, and comparison… · p. 6

B Learned repair: architecture and training

  • A repair moves an accepted, unpicked request between vehicles; the evaluator recomputes both routes and accepts only feasible moves that reduce their remaining route cost.The supervised future-score model evaluates whether an immediately improving move is likely to help the rest of the episode; its score is not a Q-learning update.Source: B Learned repair: architecture and… · p. 7
  • Paired-rollout training compares the same state and sampled future requests with and without a candidate, using greedy dispatch and workload-aware charging for up to 24 steps or 1,800 seconds.Large-scale retrieval retains 16 targets per request, and compatible repairs can be batched only when no vehicle appears in more than one move.Source: B Learned repair: architecture and… · p. 7

B.1 Architecture and training hyperparameters

The appendix identifies the deployed architecture and training/deployment settings in separate tables, with additional notes clarifying their interpretation.

B.1 Architecture and training hyperparameters

  • Table 4 lists deployed architecture hyperparameters, with MLP widths excluding the input layer and representing learned output dimensions.The proposal and future-score models have separate weights, and the PPO policy shares neither scorer’s weights.Source: Table 4 · p. 8, §B.1 · p. 8
  • Table 5 gives training and deployment hyperparameters for the reported checkpoints.Supervised checkpoints use patience-25 early stopping, making 120 a maximum rather than necessarily the selected epoch.Source: Table 5 · p. 8, §B.1 · p. 8

C PPO allocation of exact tests

The PPO selector allocates a fixed exact-test budget among cheap proposals, while a second head can decide whether to commit after exact evaluation. On held-out episodes, two-stage PPO approaches supervised repair’s cost reduction with substantially fewer exact tests.

C PPO allocation of exact tests

  • The pre-exact selector sees 16 proposals and chooses four for exact testing; a post-exact head may commit or abstain after observing the route saving.Conflicting choices are removed before batch execution, and every executed proposal must pass exact feasibility and immediate-saving checks.Source: C PPO allocation of exact tests · pp. 8–9
  • The selector is trained with clipped PPO using final episode cost, paired same-seed uniform selection as a variance-reduction baseline, and no supervised continuation label.The post-exact head is trained for eight additional episodes with the pre-exact selector frozen.Source: C PPO allocation of exact tests · pp. 8–9
  • Table 6 compares pre-exact PPO with quota-random and proposal-score controls under matched work: each retrieves 16 proposals and performs four exact tests.The comparison therefore isolates how the four-test budget is allocated; its pooled p-values are descriptive rather than confirmatory because the first six episodes were inspected before a six-episode extension.Source: Table 6 · p. 9, C PPO allocation of exact tests · p. 9
  • Two-stage PPO reduces cost by 5.23% versus no repair at the same exact work as pre-exact PPO, while staying 1.43% more costly than supervised repair and using 95.45% fewer exact tests.On the 12 held-out seeds, pre-exact PPO reduces cost by 3.18% versus no repair and performs 69.7 exact tests per episode; supervised repair performs 1,530.3 tests.Source: C PPO allocation of exact tests · p. 9

D Component ablations

The component studies test proposal ordering, future-aware commitment, computation allocation, and batching against controlled alternatives. Results support the complete repair layer, but do not establish an independent benefit from learned ordering or superiority over high-budget immediate-gain repair.

D Component ablations

  • The four direct component contrasts all have intervals crossing zero, so the experiments do not attribute the repair gain independently to learned proposal ordering.Random ordering with future-score commitment nevertheless has a positive paired interval versus no repair ([0.98, 8.93]%), with sign-flip p = 0.043.Source: D Component ablations · p. 9, Table 7 · p. 9
  • Learned future-score repair has aggregate cost 0.78% higher than the classical immediate-gain heuristic, with an effect interval crossing zero and sign-flip p = 0.998.The learned method wins 8/12 episodes, while the immediate heuristic lowers aggregate cost by 7.05% versus no repair (10/12 wins).Source: D Component ablations · pp. 9–10
  • PPO lowers cost by 5.43% versus quota-random allocation and 3.94% versus proposal-score allocation; vehicle-disjoint batching wins all ten evaluated large-scale pairs.The sparse retriever avoids constructing the large all-pairs matrix, but no feasible all-pairs control exists at n = 2000–8000, so its isolated live-EV quality effect is unclaimed.Source: D Component ablations · p. 10
  • Exact feasibility is a safety requirement: removing it would allow passenger, capacity, and battery-constraint violations and change the problem being solved.The ablation setup varies proposal order and post-test commitment while keeping the dispatcher, charger, repair epochs, candidate set, and exact-test cap fixed.Source: D Component ablations · pp. 9–10
  • Supervised repair lowers aggregate cost by 25.06% versus rolling reoptimization and serves 74.92 rather than 63.92 requests, but reoptimization is a dispatcher benchmark, not a component ablation.The paired comparison has 11/12 wins and an episode-normalized 95% interval of [13.56, 29.75]%.Source: D Component ablations · p. 10
  • In four small cases, none of the perfect-information brackets closes to proven optimality, and only the certified lower bound is rigorous for the finite-queue problem.The feasible incumbent uses a capacity-relaxed model with future requests but does not schedule shared plugs, so it is a reference rather than a queue-aware oracle.Source: D Component ablations · p. 10

E Experimental protocol

The evaluation combines held-out operational settings, paired episode-level comparisons, and additional charger-capacity, scaling, and charging studies. Large-scale results establish dynamic execution and report a batching benefit, while the distributional charging results support feasibility rather than superiority.

E Experimental protocol

  • The principal protocol uses 12 held-out Manhattan seeds with 100 requests, 30 four-seat EVs, a 1.5-hour window, and one plug per station; paired episodes are the statistical unit.The charger-capacity study repeats on the same seeds, while the two-, four-, and eight-plug extension uses four shared seeds; queue comparisons use 12 paired seeds.Source: E Experimental protocol · p. 10

F Additional operational evidence

  • The frozen supervised repair model reduces cost by 6.57% with one plug and 6.95% with unlimited plugs on the 12 held-out seeds.Four and eight plugs produce identical outcomes and no queueing in the four-seed extension.Source: F Additional operational evidence · p. 11, Table 8 · p. 11
  • Vehicle-disjoint batching improves over one repair per epoch on all ten evaluated large-scale episodes; the one-move rule accepts about 18 moves even as fleet size grows.The scale runs use n = 2000, 4000, and 8000, a frozen M = 16 retriever, and K = 96 exact cap; bounded dispatch scans at most 64 vehicles per request.Source: E Experimental protocol · p. 10, F Additional operational evidence · p. 11, Table 9 · p. 11
  • At n = 2000, 4000, and 8000, exact tests are 0.130%, 0.032%, and 0.008% of request–vehicle interactions, with peak repair memory of 0.80, 1.24, and 2.33 GB.These small post-hoc blocks show end-to-end dynamic execution but do not provide population-level estimates at these scales.Source: F Additional operational evidence · p. 11
  • Static regional decomposition lowers cost by 2.94%, 2.05%, and 0.96% at n = 1000, 2000, and 4000 while using 39.2%, 55.2%, and 67.6% fewer exact insertions than flat global search.At n = 8000, one development instance reaches a 0.08% cost gap with 68.7% fewer exact insertions and 3.16 times less solver time; this is a search-structure study, not dynamic EV control.Source: F Additional operational evidence · p. 11

G Distributional charging study

  • QPLEX compares shortlisted stations using occupancy trajectories with and without a candidate EV, adds marginal expected buffer time to travel cost, and leaves execution to the simulator’s exact FCFS queue.This is a frozen one-step distributional charging rule, not the full multi-step QDP algorithm.Source: G Distributional charging study · p. 11
  • QPLEX lowers aggregate cost by 1.77% versus nearest charging and 1.78% versus current workload, but both bootstrap intervals cross zero.Queue waiting falls 51.1% and 15.6%, respectively; the results support computational feasibility and further frozen evaluation, not a superiority claim.Source: G Distributional charging study · p. 12, Table 10 · p. 12
  • Route repair and QPLEX are not combined in the reported experiments; a reliable combination requires policy-conditioned arrival distributions and an objective valuing future vehicle availability as well as queue occupancy.Repair changes the charger-arrival process.Source: G Distributional charging study · p. 12

H Operational energy and emissions boundary

The section reports operational energy and emissions under a declared queue-idle load, alongside service and energy measurements. It distinguishes these operational findings from lifecycle or avoided-emissions claims and marks some results as development evidence.

H Operational energy and emissions boundary

  • 74.92 versus 72.75 requests were served by supervised repair and no repair, respectively; energy per served request was 0.6272 versus 0.6247 kWh, with an episode-normalized effect of −0.35% and interval [−1.58, 0.84]%.Supervised repair also used 47.08 versus 45.50 kWh and 313.89 versus 303.32 km, respectively; the passage concludes this is not a demonstrated traction-energy gain.Source: H Operational energy and emissions… · p. 12
  • Queue-idle auxiliary energy is modeled as Eaux = PauxTqueue, while operational grid accounting sums charged energy by station and time bin weighted by the declared grid factor.The section retains an April-2025 New York City-zone marginal-emissions profile for sensitivity analysis.Source: H Operational energy and emissions… · p. 12
  • At 1.5 kW, the paired bootstrap interval is [0.54, 2.98]% and sign-flip p = 0.0186; applying the loss-adjusted grid factor preserves the 1.73% operational CO2e difference.The cited emissions factor is for New York City and Westchester County.Source: H Operational energy and emissions… · p. 12

I Claim boundaries and reproducibility

Loading 2609.40225v1…