CLIPPER: Replayable Shortlisted Optimization for Repeated Spatial Coverage Planning

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.

Urban planning is a real-world spatial optimization problem that requires selecting feasible actions from large candidate spaces under practical objectives such as cost and service quality. Existing optimization and reinforcement learning methods are effective for fixed formulations, but often depend on task-specific representations and constraint handling. We propose CityPlanner, a sandbox-agent framework for executable urban planning. CityPlanner introduces UrbanSandbox, a unified file-based environment where agents inspect task files, generate plans, run evaluators, and revise decisions based on executable feedback. To make learning tractable, we further propose atomic-task reinforcement learning, which decomposes long sandbox trajectories into BuildPlan for initial construction and ImprovePlan for feedback-based refinement. Experiments on a real-world benchmark show that CityPlanner consistently outperforms heuristic, task-specific RL, and general LLM-agent baselines. Ablations verify the contributions of UrbanSandbox, atomic-task RL, and iterative deployment. We release the code and dataset at https://anonymous.4open.science/r/co-agent-C1C8

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.