CLIPPER Beyond Shortlisting: Auditable Decision Support for Changing Municipal Micromobility Policies

Julian Teusch; Jörg Philipp Müller; Monika Sester

cs.ROSep 30, 2026

CLIPPER Beyond Shortlisting: Auditable Decision Support for Changing Municipal Micromobility Policies

Authors: Julian Teusch, Jörg Philipp Müller, Monika Sester

Organizations: Institute of Computer Science Clausthal University of Technology Clausthal-Zellerfeld, Germany · Institute of Cartography and Geoinformatics Leibniz University Hannover Hannover, Germany

Abstract

In municipal planning workshops, planners and other stakeholders compare shared-micromobility parking policies by varying no-parking zones, retained sites, spacing, or area allocations. Each edit changes feasible sites and how much demand they cover, so the alternative must be reoptimized on the same spatial data. Full-set greedy, the transparent reference for this task, takes tens of seconds per alternative at city scale. We present Constraint-exact Low-latency Iterative Planning with Pooled Evaluation and Replay (CLIPPER), an optimizer with audit functions developed for requirements elicited with the City of Braunschweig. In each greedy round, it forms a deterministic candidate pool of bounded size, computes how much still-uncovered demand each candidate would add, and rejects candidates that violate an active constraint. An optional offline audit scans every remaining feasible candidate and records what the restricted pool omitted. We evaluate these functions on complete eleven-state edit chains ( $E_0,\ldots,E_{10}$ ) in Braunschweig, Munich, and Berlin. With $K=1024$ candidates per group, the fixed-width mode CLIPPER-F has mean coverage gaps to full-set greedy under the same policy of 0.245, 0.003, and 0.001 percentage points in Braunschweig, Munich, and Berlin, respectively, while mean rollout time falls by factors of 13.6--28.9; no audited run terminates while a candidate outside the pool could still increase coverage. Plans computed from two checksummed versions of Braunschweig's official no-parking-zone data differ in 30 of about 540 selected sites although coverage moves by only about 0.1 percentage points. These changes still require municipal assessment and implementation. The findings inform a proposed municipal process that versions policy inputs, reports site changes beside coverage, and scans the full candidate set before a final decision.

Figures & tables

Panel of Figure 1 — Figure 1 . CLIPPER pipeline. Scenario constraints govern both the candidate pool and greedy selection. The screening rule (defined later in this section) can request a larger pool or an offline audit. That audit scans all remaining candidates and calculates the coverage gain omitted from the restricted pool. CLIPPER pipeline diagram: scenario constraints feed per-group candidate rules that form a pooled candidate set, a restricted-pool greedy selector selects feasible actions under constraint checks, and an optional screening rule can request candidate-budget expansion.

Panel of Figure 2 — Figure 2 . Illustrative Braunschweig diagnostic on a separate $25\text{\,}\mathrm{m}$ grid; quantitative comparisons use the $10\text{\,}\mathrm{m}$ support in Section 5.1 . Panel A shows demand and candidates; B marks baseline sites in teal; C shows selected sites (blue), locks (dark squares), and exclusions (coral); D marks retained, removed, and added sites in gray, coral, and teal. At $E_{5}$ , $162$ candidates are forbidden, $90$ sites are locked, and $25\text{\,}\mathrm{m}$ spacing is active; the plan retains $443$ baseline sites and adds $157$ replacements. Four maps show demand and candidate support, the baseline selected plan, the edited plan with exclusions, locks, and spacing, and retained, removed, and added sites.

City	Trips	Cand.	Budget	Groups	F slots
Braunschweig	$243\,649$	$60\,495$	600	29	$29\,696$
Munich	$4\,958\,963$	$32\,907$	1200	16	$16\,384$
Berlin	$12\,306\,947$	$68\,922$	1600	36	$36\,864$

Table 1 . Full-data instances. F slots are the maximum CLIPPER-F(

K=1024

) entries per round before filtering and deduplication; CLIPPER-A uses

8192

total slots.

State	$E_{0}$	$E_{1}$	$E_{2}$	$E_{3}$	$E_{4}$	$E_{5}$	$E_{6}$	$E_{7}$	$E_{8}$	$E_{9}$	$E_{10}$
Baseline sites inside $75\text{\,}\mathrm{m}$ exclusions (%)	0.0	2.5	5.0	5.0	7.5	7.5	10.0	12.5	12.5	15.0	17.5
Baseline sites kept as locks (%)	0.0	0.0	0.0	10.0	10.0	15.0	20.0	20.0	25.0	30.0	35.0
$150\text{\,}\mathrm{m}$ exclusion halos	0	0	0	0	0	0	0	0	10	15	20
Minimum network spacing (m)	0	0	0	0	0	25	35	45	50	55	60

Table 2 . Complete configuration of the controlled edit chain. Exclusions target a growing prefix of baseline sites ordered by individual coverage contribution; locks use contribution and spatial dispersion. Each halo is an exclusion zone centered on a dense endpoint cell.

Balanced caps: full-set greedy and CLIPPER-F
Method / control	Mean coverage (%)			Mean gap to control (pp)			Mean rollout (s)			Rollout/control
	BS	MUC	BER	BS	MUC	BER	BS	MUC	BER	BS	MUC	BER
Full-set greedy	90.4	69.7	59.5	0.000	0.000	0.000	24.029	22.943	52.685	1.000	1.000	1.000
CLIPPER-F(1024)	90.1	69.7	59.5	0.245	0.003	0.001	1.762	1.493	1.825	0.073	0.065	0.035
Coverage-prioritized caps: full-set greedy and CLIPPER-A
Full-set greedy	96.8	81.2	69.5	0.000	0.000	0.000	28.343	22.931	48.826	1.000	1.000	1.000

Table 3 . Main comparison on complete data, averaged over

E_{0},\ldots,E_{10}

. BS, MUC, and BER denote the three cities; pp denotes percentage points. Each block pairs CLIPPER with full-set greedy under the same caps, on the same host and with the same rollout timing boundary ( Section 5.1 ). Gap is control minus method coverage; rollout/control is their time ratio.

(a) Audit of a candidate order computed once per state
City	$K$	Mean coverage (%)	Audited rounds: no larger feasible gain omitted (%)	Omitted gains / added coverage	Runs stopped by pool limit (of 11)
Braunschweig	256	77.248	4.4	3.295	11/11
	512	85.677	13.3	1.071	11/11
	1024	90.132	51.0	0.140	0/11
Munich	256	64.623	10.5	0.748	6/11
	512	69.395	66.4	0.046	0/11

Table 4 . Results on

E_{0},\ldots,E_{10}

. Panel (a) scans the full remaining set after each selection and at termination; it reports rounds with no larger feasible gain omitted, normalized omitted gains, and stops caused by the pool limit. Panel (b) uses the stepwise executor from Section 5.1 ; its timer includes initialization, updates, heap maintenance, pool construction, and selection. The host is shared with Table 3 , but the execution path and timer differ; panel (b)’s times are compared only within that panel. Coverage values are means; wall times and paired per-state time ratios are medians. The final column normalizes post-initialization score updates by the product of selection calls and stored candidate rows.

(a) Policy resources and resulting coverage
Resource date	Zones	Area ( $\mathrm{km}^{2}$ )	Forbidden cand.	Fixed scores: cov. (%)	Updated scores: cov. (%)
2025-01-28	51	6.275	$1435$	91.609	91.807
2026-02-17	55	9.429	$4653$	91.515	91.701

Table 5 . Braunschweig case with two official no-parking-zone resources and all

60\,495

candidates. Both runs use

K=1024

, identical demand, a 600-site budget, and no locks or spacing; of 30 removed sites, 22 became forbidden and eight remained feasible.

Stage	Inputs, outputs, and acceptance check
Register	Input: policy geometry, candidates, demand, coordinate system, validity date. Check: source checksum, valid geometry, and candidate intersections.
Calibrate	Input: representative recorded states and local time/quality tolerances. Check: choose $K$ , cap policy, and candidate rule; meet local targets and scan for improving candidates outside the pool when a run stops.
Compare	Input: exclusions, locks, spacing, caps, and facility budget for each alternative. Check: feasible plan, coverage, changed sites, terminal status, and input/output checksums.
Final decision	Input: chosen policy state and audit level. Output and check: scan all remaining candidates, archive inputs and outputs, and obtain policy-owner sign-off.

Table 6 . Municipal pilot pathway. The municipality owns policy and acceptance thresholds; CLIPPER supplies plans, comparisons, and audit records.

Explore similar work

Feb 8, 2026cs.RO

Multi-Agent Route Planning as a QUBO Problem

Multi-Agent Route Planning considers selecting vehicles, each associated with a single predefined route, such that route-level coverage utility is maximized while redundant spatial overlaps are limited. This paper gives a formal problem definition, proves NP-hardness by reduction from the Weighted Set Packing problem, and derives a Quadratic Unconstrained Binary Optimization formulation whose coefficients directly encode route utility rewards and pairwise overlap penalties. A single penalty parameter

λ

controls the coverage--overlap trade-off. We distinguish between a soft regime, which supports multi-objective exploration, and a hard regime, in which the penalty is strong enough to effectively enforce near-disjoint routes. We describe a practical pipeline for generating city instances, constructing candidate routes, building the QUBO matrix, and solving it with a binary quadratic programming baseline (Gurobi), simulated annealing, and D-Wave hybrid quantum annealing. Experiments on Barcelona instances with up to

10{,}000

vehicles reveal a clear coverage--overlap knee and show that Pareto-optimal solutions are mainly obtained under the hard-penalty regime, while D-Wave hybrid solvers and Gurobi achieve very similar objective values on matching configurations with only minor runtime differences as problem size grows.

Renáta Rusnáková, Martin Chovanec, Juraj Gazda

Department of Computers and Informatics, Faculty of Electrical Engineering and Informatics, Technical University of Košice, Letná 9, 042 00 Košice

Vehicle Routing Problem Multi-Agent Path Finding Quadratic Programs Quantum Annealing

Ask PDF

Jul 17, 2026cs.LG

From Feasibility to Desirability: Plan, Learn, Adapt (PLA) Framework for Personalized On-Device Itinerary Generation

Generating personalized trip itineraries is a complex planning task and involves a tension between hard combinatorial feasibility and soft latent desirability. Classical optimization enforces constraints but fails to capture subjective traveler preferences. While learning-based approaches model preferences, they cannot guarantee feasibility. Mobile deployment imposes additional resource constraints on both. To address this, we propose Plan, Learn, Adapt (PLA), a three-stage framework for personalized on-device itinerary generation. The Plan stage builds a heterogeneous ensemble of lightweight planners that produces structurally diverse feasible candidates. From pairwise itinerary comparisons, Learn fits a compact Bradley-Terry reward model that captures emergent schedule properties such as pacing, geographic coherence, and day balance, which per-POI signals miss. Finally, Adapt applies feasibility-preserving local refinement within a device-aware compute budget; every intermediate state is feasible by construction. On 2,519 pairwise human comparisons across more than 100 U.S. cities, the reward-guided ensemble achieves a 67.8% win rate, 11.2 percentage points above the best single planner, with 100% feasibility. Three frontier LLMs, GPT-5, Claude Opus 4.5, and Gemini 3 Pro, achieve 0% feasibility under the same constraints. The reward model generalizes across held-out cities, with a 67.6% mean leave-one-city-out accuracy. In production deployment within FlyEnJoy, PLA increased itinerary completion rates by 91%, with 109.9 ms average on-device latency.

Himel Dev, Tanmoy Sen, Madhusudan Basak +1

Tech LLC · University of Virginia · University of Massachusetts Amherst

Journey Personalization Emergence

Ask PDF

Aug 1, 2026cs.LG

Paris as a 15-Minute City: An Explainable AI Perspective

The 15-minute city promotes access to everyday services within a short walk or bicycle ride, but its relationship with observed mobility remains difficult to quantify. We investigate this relationship in the Paris metropolitan area using mobility trajectories from the NetMob 2025 Data Challenge, enriched with INSEE sociodemographic data and OpenStreetMap points of interest (POIs), yielding approximately 70,000 trip segments after stop-based segmentation and data cleaning. We construct walking- and cycling-based indicators of local service availability and examine their associations with trip duration, transport mode, and short-trip car use. Higher POI availability is associated with less private motorized travel and more active mobility, although this relationship is substantially weaker in the outer agglomeration. Gradient-boosted tree models interpreted with explainable machine-learning methods consistently identify trip purpose, home--work distance, local service availability, vehicle ownership, public-transport subscription, and sociodemographic context as important predictors. For short trips, high POI density is associated with lower car use, while car ownership and driving-licence availability are associated with higher predicted car use; where services are sparse, public-transport subscription is associated with lower predicted car dependence. Finally, explainable AI (XAI) methods are used to examine how feature attributions change under alternative assumed variable orderings. The results are consistent with central assumptions of the 15-minute city while revealing substantial spatial and demographic heterogeneity. They also demonstrate how explainable machine-learning methods can complement accessibility indicators and identify locally relevant hypotheses for urban-mobility policy.

András J. Molnár, Csaba I. Sidló, Rita Rónai +1

Institute for Computer Science and Control, Hungarian Research Network, Hungary

Urban Mobility Urban Environments Bivariate Bicycle Codes Explainable Artificial Intelligence

Ask PDF