Scheduling Buses for the Wait Tolerances Riders Actually Have

Some bus scheduling methods try to maximize the number of riders who wait less than a fixed tolerance $\theta$, for example ten minutes, instead of minimizing average wait. On Singapore's network this doubled, and in some scenarios tripled, the number of satisfied passengers compared with fixed-interval timetables (Towards an Optimal Bus Frequency Scheduling: When the Waiting Time Matters). The whole method rests on one number, and nobody measures it well. Agencies take $\theta$ from what riders say in surveys. Matching Danish travel diaries to smart card taps shows that people misreport their own trip times by a median of 11.34 minutes, with an interquartile range of 28.14 minutes, and that the error depends on interview mode, trip order and schedule type (Analyzing the reporting error of public transport trips in the Danish national travel survey using smart card data). A threshold built from remembered minutes and then applied to clock minutes can be off by about as much as the threshold itself.

This note proposes treating wait tolerance as a distribution across riders rather than a single number. The distribution would be estimated from what riders do and from survey answers corrected for measurement error, and then used to schedule buses. The proposal has three parts. First, a scheduling objective that counts each rider's probability of being satisfied. Second, a way to read tolerance off smart card data using transfer passengers, whose arrival at the stop is timestamped by their tap-off. Third, a sensitivity test that comes first and could end the project: if schedules from the existing method barely change as $\theta$ moves across its plausible range, then self-report bias does not matter here, and we will say so.

Core Claim

The satisfaction-boosting problem (called FAST) counts a passenger as served if some dispatched bus picks them up within $\theta$ (Bus Frequency Optimization: When Waiting Time Matters in User Satisfaction). We replace the hard cutoff with each rider's probability of tolerating the wait they get:

$$f(S) = \sum_{p} G_p\Big(\min_{b \in S} w_{p,b}\Big),$$

where $S$ is the set of dispatched buses, $w_{p,b}$ is the wait passenger $p$ would face if bus $b$ served them, and $G_p(w)=\Pr(\theta_p \ge w)$ is the probability that this passenger's tolerance is at least $w$.

Two facts make this practical. Because $G_p$ is decreasing, each passenger's term equals the largest $G_p(w_{p,b})$ over dispatched buses. That makes $f$ a weighted facility-location function, which is monotone and submodular, meaning each extra bus adds less than the one before. So the $(1-1/e)$ guarantee of plain greedy carries over unchanged. The partition step of PartGreedy and ProPartGreedy should also carry over, since it only uses the fact that clusters of routes rarely share passengers; we will verify this rather than assume it. The original FAST objective is the special case where $G_p$ is a step function at a common $\theta$.

The claim that would be new is empirical. When $G_p$ is estimated from behavior, schedules optimized against it serve measurably more riders within their actual tolerance than schedules optimized against a survey-derived scalar, a corrected scalar, or a scalar tuned on held-out behavior. The gain should be concentrated on low-frequency peripheral routes, where a mis-set cutoff decides whether a bus is dispatched at all.

Research Program

The algorithm. We reimplement Greedy, PartGreedy and ProPartGreedy on the Singapore bus testbed used by the FAST papers and by the FASTS: A Satisfaction-Boosting Bus Scheduling Assistant. The testbed has 396 routes, 28 million touch-on/touch-off records from one week in April 2016, and 451,000 candidate one-minute departures. We then swap in the probabilistic objective. The changes are small: the forward and inverted indexes stay the same, and a marginal gain becomes a sum of probability improvements instead of a count. We keep the original papers' omission of vehicle capacity, so the comparison isolates the threshold. Capacity is a separate, known gap and we leave it out on purpose.

Revealed tolerance. Bus smart card data normally hides when a rider reached the stop, which is why wait tolerance cannot be read off directly. Transfer passengers are the exception. Their tap-off from the first bus timestamps their arrival at the transfer stop. When several routes serve their onward destination, two choices reveal how long they are willing to wait: boarding an early but slower bus instead of waiting for a later, faster one, and walking to a different stop. This borrows the logic of the Singapore metro study, which inferred the cost of standing from riders choosing to travel backwards (Valuation of sitting and standing in metro trains using revealed preferences). It also borrows the Shenzhen finding that riders let trains pass for comfort (Estimation of Passenger Route Choice Pattern Using Smart Card Data for Complex Metro Systems). In the transfer setting, the rider's wait has a known start time, which those metro studies lacked.

We fit a choice model in which the probability of abandoning the wait for the faster bus rises once elapsed wait passes a latent tolerance. Tolerance varies by time of day, route headway class and, where the card type is recorded, rider group. The fitted model gives $G_p$. One limitation is that transfer riders may tolerate waits differently from riders starting a trip. We test this by checking whether the estimated distribution, applied to riders starting a trip, predicts the observed drop in boardings on routes whose headways sit just above versus just below the estimated tolerance quantiles.

Survey prior with error correction. Agencies will keep using surveys, so the second estimator starts from a stated-tolerance distribution and deconvolves it. Deconvolution here means removing the known spread of reporting error to recover the underlying distribution. We use an error model whose shape and moderators are taken from the Danish matching study: interview mode, trip order, and fixed versus flexible schedules. The assumption that stated tolerances carry the same error structure as reported trip times is the weakest link in the proposal, so we treat it as a hypothesis. The revealed estimate decides it. If the corrected survey distribution lands near the revealed one, agencies have a cheap route to good thresholds. If it does not, the survey route should be dropped.

These methods play different roles. The Danish matching study motivates the work but never connects its error distributions to any operational decision. The FAST papers supply the optimization method and are the thing we modify. The revealed-preference metro studies supply the identification strategy. We do not use stated-preference choice experiments as the source of tolerances. The intention–behavior evidence summarized in the file, for example stated support for contact-tracing apps that did not show up in actual adoption (Samuel Altmann et al.), argues against relying on them.

Baselines

Naive reference points. The do-nothing reference is the fixed-interval timetable used in the FAST papers. Below it sits average-wait minimization, which those papers show starves peripheral routes. The strongest existing method is ProPartGreedy with a single scalar $\theta$. The reported results on the Singapore testbed come from Towards an Optimal Bus Frequency Scheduling: When the Waiting Time Matters:

  • it "doubled (and in some scenarios tripled)" satisfied passengers relative to fixed intervals;
  • it finished within 9.4% of global Greedy;
  • it ran roughly ten times faster than Greedy.

Scalar-threshold variants. We run ProPartGreedy with a scalar $\theta$ chosen in three ways, all on the same week of data, the same candidate bus set, the same bus budget and the same $\epsilon$:

  • the raw survey value;
  • the survey value shifted by a flat correction derived from the Danish median error;
  • the scalar that maximizes expected satisfaction on a held-out revealed-tolerance split.

The third is the strongest competitor. It uses the same behavioral data as our method but compresses it into one number. It gets the same tuning budget we give to estimating the distribution.

Ablations. Two ablations isolate the key ingredient. One uses a single pooled distribution shared by all riders, which tests whether rider-level differences matter or only the spread. The other uses the uncorrected survey distribution, which tests whether the error correction does any work.

Evaluation and the bar to clear. All schedules are scored on held-out days by expected satisfied passengers under the revealed $G_p$, plus the satisfied-rider count at several fixed cutoffs, so the result does not depend on our own estimate. Beating the tuned scalar means a consistent gain at equal bus budget across held-out days, larger than day-to-day variation, and concentrated on peripheral routes. Matching it would mean the distribution adds nothing beyond choosing one good number.

Why not just…?

Why not just subtract a flat correction from the survey threshold? The Danish errors are wide and depend heavily on context: web respondents had a median error of 7.23 minutes against 12.72 for telephone, and the interquartile range was 28.14 minutes. A single shift moves the cutoff but cannot represent the spread. The flat-correction baseline above measures this directly. If it comes close to our method, the honest conclusion is that only the location of the threshold matters.

Why not just tune one scalar $\theta$ on behavioral data? This is the real threat. If tolerance is tightly clustered, or if the greedy schedule is insensitive to the spread, a well-chosen scalar captures almost everything. The facility-location form predicts an advantage only where the tolerance distribution is wide relative to route headways. We will report the gain separately by headway class. If it disappears everywhere, the contribution reduces to "estimate $\theta$ from transfer behavior," which is still useful but much smaller.

Why not just optimize against the worst case over a range of thresholds? A max-min schedule protects against misspecification. It does so by optimizing for the least tolerant plausible rider, which wastes buses on corridors where riders are patient. It also gives up the plain submodular structure that the greedy guarantee needs. We include it as an extra comparison on a subset of the network where the cost of solving it is manageable.

Proposed Experiments

The first experiment decides whether the project continues. On the Singapore testbed we run ProPartGreedy at $\theta$ values spanning the survey-plus-error range implied by the Danish interquartile range. For each, we measure two things: how much the dispatch set overlaps with the dispatch set at the central $\theta$, and the loss in satisfaction when a schedule built for one $\theta$ is scored at another. If the overlap stays high and cross-scored losses are a few percent, self-report error in $\theta$ is operationally harmless. The project would then stop and report that negative result.

If the losses are substantial, we estimate revealed tolerance from transfer passengers. We then check it on riders starting a trip using the headway-boundary test described above, and run the full baseline comparison on held-out days.

The final check asks whether the error-corrected survey distribution matches the revealed one closely enough that schedules built on each perform within the gap separating the tuned scalar from the full method. If it does, agencies can use cheap surveys with a correction; if not, surveys should not set scheduling thresholds at all.