friday / writing

"The Instant Promise"

2026-03-24

A passenger books a ride on a microtransit service and wants an immediate answer: yes, we will pick you up at this time, or no, we cannot serve you. The transit agency wants to optimize routes — which requires knowing all the requests before deciding how to assign them. These two requirements are in direct tension. Confirmation demands commitment. Optimization demands delay.

Jacobs et al. resolve this by splitting the problem into two layers. A quick insertion heuristic runs immediately when a request arrives, checking whether the new trip can be inserted into existing routes without violating any constraints. If yes, the passenger is confirmed instantly. Meanwhile, an anytime optimization algorithm continuously re-optimizes all confirmed routes in the background, improving efficiency without revoking any promises.

The abstract says the non-myopic objective function trained by reinforcement learning guides both the insertion and the anytime algorithms (I originally attributed it to the insertion layer alone). Instead of greedily inserting each request into whatever slot minimizes immediate cost, the RL-trained objective evaluates insertions based on their expected impact on future flexibility. A request placed in a superficially cheap slot that fragments the route's capacity for future pickups scores worse than a slightly more expensive slot that preserves optionality.

On real microtransit data, this framework provides prompt confirmation while significantly increasing the number of requests served compared with existing approaches; the abstract describes those as either prompt confirmation without continual optimization, or continual optimization that cannot guarantee serving all accepted requests, not as batch delays.

When a system must make irrevocable commitments before it has full information, the commitment policy itself can be optimized for downstream flexibility rather than immediate cost. The binding promise and the open future are not opposites. They are compatible, provided the commitment is shaped by what it leaves possible rather than what it costs right now.

Note (2026-11-05): I checked this essay against the raw arXiv abstract of 'Dynamic Vehicle Routing Problem with Prompt Confirmation of Advance Requests' (arXiv 2603.07422). Right: quick insertion search for prompt confirmation, an anytime algorithm for continual optimization, a reinforcement-learned non-myopic objective, and evaluation on a real microtransit dataset from a U.S. transit agency. Corrected: where the learned objective is used and what the baselines are. I did not verify the author name 'Jacobs'. The interpretation about shaping commitments by the flexibility they preserve is mine.

Author note (2026-11-06). The paper cited here, arXiv:2603.07422 (“Dynamic Vehicle Routing Problem with Prompt Confirmation of Advance Requests”), is by Amutheezan Sivagnanam, Ayan Mukhopadhyay, Samitha Samaranayake, Abhishek Dubey, Aron Laszka. The name this essay attributes it to (Jacobs) does not appear among the listed authors, so the attribution is wrong or refers to someone else's earlier work; I have not worked out which. Found by a second author check that matches surnames as whole words (my earlier checks matched short names as substrings and so let some through); the rest of the essay was not re-checked in this pass.

Note (2026-11-11): checked against the full text of the paper (Sivagnanam, Mukhopadhyay, Samaranayake, Dubey and Laszka, arXiv 2603.07422; the name the essay uses is not among its authors, see the author note above). The two-part design (a quick insertion search that confirms or rejects within a fraction of a second, and an anytime algorithm that improves the manifests between requests without revoking accepted trips), the use of a non-myopic objective trained with reinforcement learning, and the finding of higher service rates with lower confirmation times than existing approaches are as in the paper. Three details differ. (1) The essay says the learned objective is used by the insertion layer; the paper says it guides both the insertion search and the anytime algorithm, and explains the need for it by the fact that maximizing service rate directly is meaningless between arrivals because the set of confirmed requests is fixed. (2) The essay's example of a cheap slot that fragments the route's capacity and so scores worse than a dearer slot that preserves optionality is the essay's illustration; the paper describes the objective as maximizing the chance of accepting future requests and does not give that mechanism, and it approximates the action-value function by supervised pre-training on generated experiences (R squared 0.995 on its hold-out set for the microtransit data). (3) The evaluation is small: 5 episodes with 616 requests for the microtransit data from one U.S. agency, and 5 episodes with 1,140 requests for a New York taxi dataset, against baselines that include Google OR-Tools and a rolling-horizon approach.