Repairing Stale Transit Networks: Targeted, Persistence-Aware Re-optimization Under Drifting Demand

A city's bus network is designed against a snapshot of the world: which roads exist, how long each link takes to drive, and how many people travel between each pair of neighborhoods. That snapshot goes out of date. A road closes for two years, a new office park slowly pulls commuters into a corridor, a stadium creates a surge every other weekend. Today's learned transit designers treat every new snapshot as a fresh problem. They either re-optimize the whole network or keep the old one. This proposal treats an optimized network the way recent language-model work treats stored facts: as knowledge with a dependency structure and a shelf life. When inputs change, we find which routes those changes actually made obsolete, repair only those routes, and discount changes that are unlikely to last long enough to justify moving a bus line.

The two-sentence version is this. Most of a well-designed network stays good after a demand or road shift, and the damage concentrates in the few routes that carry the changed trips. If we trace each route back to the trips and road links it depends on, and weight each change by how long it is likely to persist, a neural route-editing policy should recover near-optimal trade-offs between passenger time and operator cost far faster than re-running the search. It should also change far fewer routes, which matters because route changes confuse riders.

Core Claim

Transit network design is a multi-objective problem: shorter passenger trips cost more vehicle-hours, and the goal is the set of designs where neither can improve without hurting the other (the Pareto front). The strongest learned method here is the Neural Evolutionary Algorithm (NEA), an evolutionary search whose random mutations are replaced by a graph attention network trained with reinforcement learning to propose new routes (Learning Heuristics for Transit Network Design and Improvement with Deep Reinforcement Learning; A Neural-Evolutionary Algorithm for Autonomous Transit Network Design). It supplies our search machinery. Its authors note that it assumes static demand matrices and fixed travel times, and that its moves replace whole routes chosen without regard to where the cost came from. We claim that in a sequential setting, where demand and link times drift over a series of planning epochs, two ingredients from the knowledge-obsolescence literature change what good re-optimization looks like.

The first ingredient is dependency-based invalidation. Can Agent Memory Systems Track Evolving State? shows that agents anchor on superseded values even when the current ones are in front of them. Its StateMem system fixes this by tagging stored facts with what they were derived from and deterministically invalidating dependents when a base value changes. That work motivates the mechanism and supplies its design. We apply the same structure to networks. Each route is linked to the origin-destination trips whose best transit path uses it and to the road links it traverses. When an input changes, the routes downstream of the change are marked stale and ranked by how much of the cost increase passes through them.

The second ingredient is persistence weighting. Mitigating Temporal Misalignment by Discarding Outdated Facts predicts how long a fact will remain true and discounts confidence accordingly. We transplant that idea from text facts to demand changes. Let $D_{ij}$ be the current demand between zones $i$ and $j$, $\Delta_{ij}^{k}$ the change in that demand caused by event $k$, and $p_k(H)$ the predicted probability that event $k$ is still in effect at the end of the planning horizon $H$. Repair then optimizes against

$$\tilde D_{ij} = D_{ij}^{\text{old}} + \sum_k p_k(H)\,\Delta_{ij}^{k},$$

so the network chases a weekend surge only to the extent the surge is expected to last. Link travel times are treated the same way. Without this ingredient, a repair method faces the transit version of the "distraction susceptibility" that Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams documents for language models under frequent updates: it over-reacts to transient fluctuations and pays for that in route churn.

What is new is the combination: localized, dependency-traced repair of a Pareto set of transit networks, steered by predicted persistence, and judged on a third objective (stability) that static design ignores. Nothing in the file connects these communities. The obsolescence papers all live in text QA, dialogue, or narrative streams, and the transit papers all assume stationary inputs.

Research Program

We first build a drifting-demand benchmark, because none exists. The base instances are Mumford3 (127 nodes) and the Laval network (632 nodes, over 548,000 trips) used by the NEA papers. A scripted event generator perturbs them over twelve planning epochs. Event types include persistent link closures, temporary closures, gradual corridor growth, recurring event surges, and one-off surges. Each event carries a ground-truth duration. Because the generator knows true durations, a persistence predictor trained on its outputs could look artificially good. We therefore evaluate with three predictors: an oracle, a learned classifier over event features and recent history (the analog of Zhang and Choi's duration classes), and a deliberately miscalibrated version. This tells us how much accuracy the method actually needs. As an external sanity check, final networks for a Chattanooga instance are run through Transit-Gym: A Simulation and Evaluation Engine for Analysis of Bus Transit Systems, which simulates bus operations microscopically on its 2014 demand matrix. This confirms that gains in the analytic cost model do not vanish under simulated traffic. One limitation is that the file contains no real longitudinal origin-destination data, so the drift is synthetic. We say so plainly and vary the event mix widely rather than claim realism.

The method itself has three parts. A dependency index maps every route in every archived network to the trips and links it serves; it is rebuilt cheaply after each shortest-path evaluation. An attribution step scores each route by the cost increase on the trips that depend on it under $\tilde D$. A repair loop runs the NEA mutator but restricts its proposals to the top-ranked stale routes. NEA's policy was trained on network construction and then deployed on improvement, a mismatch its authors flag. Repair is a third task, so we also fine-tune the policy on repair episodes and report both the frozen and fine-tuned versions. The objectives are passenger cost, operator cost, and churn, measured as the number of route stops added or removed relative to the previous epoch's deployed network.

Baselines

The naive reference point is to do nothing: keep the epoch-zero Pareto archive and deploy from it unchanged. This measures how much drift actually hurts. If the cost of inaction is small on our event mix, the whole question is moot, and we will report that.

The quality ceiling is NEA re-run from scratch at every epoch on the true current demand, with the tuning budget its authors use. On static Mumford3, NEA is reported to beat a standard evolutionary algorithm by 33% to 53% depending on cost weighting (A Neural-Evolutionary Algorithm for Autonomous Transit Network Design), to improve on previously best-known operator costs by up to 4.8%, and to cut passenger cost by 30% and operating cost by 19% on Laval relative to the existing network (Learning Heuristics for Transit Network Design and Improvement with Deep Reinforcement Learning). There is no reported result found on Lacuna for any transit design method under drifting demand, so all dynamic numbers will be new.

The strongest competitor is warm-started NEA. This is the standard dynamic-optimization move: seed each epoch's search with the previous archive and run untargeted mutation on current demand. A related precedent is Augmenting Transit Network Design Algorithms with Deep Learning, where neural warm starts reached a standard hyper-heuristic's final score at 2% of its steps on Mumford3. Warm-started NEA gets the same evaluation budget, the same policy weights, and the same hyperparameter search as our method.

We count it as beaten if our method matches warm-started NEA's hypervolume (the volume of objective space dominated by the archive) within 1% at no more than a quarter of its mutation steps, or reaches higher hypervolume at equal steps. In either case it must also show lower cumulative churn and lower cumulative passenger-plus-operator cost summed over all epochs. We also include scenario-robust NEA, which optimizes the average cost over sampled future demand scenarios drawn from the same event generator.

Three ablations isolate the ingredients. The first is targeting with raw rather than persistence-weighted demand. The second is persistence weighting without targeting. The third is random targeting: the mutator restricted to the same number of routes, chosen uniformly. Random targeting is the key control. If it matches dependency-based targeting, the gain comes from making smaller edits, not from knowing which routes are stale.

Why not just…?

Why not just warm-start NEA from the last archive? This is the real threat, and we do not claim to know it loses. NEA's learned mutator might already focus on high-cost routes, which would make explicit dependency tracing redundant. The random-targeting and warm-start comparisons settle this directly. If warm-started NEA matches us on hypervolume per step, the remaining claim is the persistence and churn result, and we would rewrite the paper around that.

Why not just re-evaluate the old archive under new demand and keep the non-dominated designs? Evaluation is cheap; it amounts to shortest paths on the transit graph. This filter is an excellent detector of which designs got worse, and we use it as the first stage of every method. It cannot produce new designs, though, so under persistent closures it should lose badly, while under transient events it may be nearly optimal. The epoch-by-epoch breakdown by event type will show where this boundary lies.

Why not just optimize robustly over demand scenarios from the start? If future events were well characterized, a robust design could make repair unnecessary. That assumes the event distribution is known at design time, which is exactly the stationarity assumption the obsolescence literature warns against. SituatedQA: Incorporating Extra-Linguistic Contexts into QA found models 15 points less accurate on facts that changed than on stable ones. We test this by running scenario-robust NEA when the generator's event mix shifts mid-sequence. We expect it to hold up under the trained mix and degrade after the shift.

Why not put a language-model agent with retrieval over live feeds in charge? The inputs here are structured matrices, not documents, so retrieval is not the bottleneck. Both Can Agent Memory Systems Track Evolving State? and the OAKS benchmark report that free-form reasoning over update streams anchors on stale values or over-reacts to noise. We borrow those papers' structural fix, not the agent.

Decisive Checks

Three results would decide the proposal. On Mumford3 and Laval, dependency-targeted repair must beat random targeting at equal step budgets; if it does not, the invalidation idea fails. Persistence weighting must cut cumulative churn on transient-heavy event mixes without raising cumulative cost on persistent-heavy ones, and this must hold with the learned predictor, not just the oracle. Transit-Gym simulation of the final Chattanooga networks must preserve the ranking among methods. If all three hold, the result is a concrete demonstration that the bookkeeping developed for stale language-model knowledge transfers to the maintenance of physical infrastructure plans.