A delivery fleet has a hard emission cap. Every route burns fuel. Every delivery creates value. The constraint is absolute: once the cap is reached, the fleet stops. The question is not which deliveries to make but which deliveries to refuse.
Tassel et al. formalize this as the Dynamic and Stochastic Vehicle Routing Problem with Emission Quota and show that reinforcement learning learns to reject profitable deliveries early in the day to preserve emission budget for later ones. The optimal policy says “no” to good orders now so it can say “yes” to better orders — or simply more orders — later. A myopic policy that accepts every profitable delivery until the cap binds will serve fewer total customers than one that strategically refuses from the start.
The mechanism is temporal asymmetry under a hard constraint. Early in the day, the fleet has budget but uncertainty about future demand. Late in the day, the fleet has information but may have exhausted its budget. A greedy policy converts budget to revenue as fast as possible, which means the late-day information arrives after the budget is gone. The RL agent learns the opposite: conserve budget through the uncertain period so that late-arriving, high-information decisions can still be acted on.
The abstract says the two-layer framework is designed to facilitate anticipatory rejections of demands and generation of new routes, using hybrid algorithms that combine reinforcement learning with combinatorial optimization, and that the approach is relevant for different inputs even when the horizon is uncertain. My description of which orders are rejected, and of the rejections as surgical rather than proportional, is not in the abstract.
Under a hard cumulative constraint, the value of a resource is not its current yield but its option value — the future decisions it keeps available. Accepting a good opportunity that consumes a shared resource can be strictly worse than refusing it, not because the opportunity is bad, but because accepting it forecloses better uses of the resource that have not yet appeared.
Note (2026-11-05): I checked this essay against the raw arXiv abstract of 'Demand Acceptance using Reinforcement Learning for Dynamic Vehicle Routing Problem with Emission Quota' (arXiv 2603.13279). Right: the problem DS-QVRP-RR with a global emission constraint and anticipatory rejections. Not in the abstract: that a myopic policy serves fewer total customers, the temporal-asymmetry mechanism, and the claim that the learned policy rejects profitable early orders. Those are my reading of the setup, not reported findings; the abstract compares against traditional methods without giving numbers. The option-value moral is mine. I did not verify the author name 'Tassel'.
Author note (2026-11-06). The paper cited here, arXiv:2603.13279 (“Demand Acceptance using Reinforcement Learning for Dynamic Vehicle Routing Problem with Emission Quota”), is by Farid Najar, Dominique Barth, Yann Strozecki. The name this essay attributes it to (Tassel) 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): full-text check of arXiv 2603.13279 (Najar, Barth and Strozecki, AAAI 2026). That a reinforcement-learning policy can rationally reject some demands early to leave room under a hard emission quota is as in the paper, which compares it with a first-arrived-first-served heuristic and a multiple-scenario sampling method and reports substantial gains. Four points differ. (1) The paper's objective is to maximize the number of destinations served; the essay writes of 'profitable' deliveries, 'good orders' and orders' value. (2) In the paper, accept or reject decisions are made online while requests arrive during the planning phase, before the tour begins, and the tour is executed after all requests have arrived; the essay's picture of a fleet that burns its emission budget during the day, so that 'late-day information arrives after the budget is gone', is the essay's own framing. (3) The essay says the rejections are 'surgical': orders with high emission cost relative to value or that need inefficient detours. I did not find that in the paper; its discussion says the learned model, using engineered state features, learns to discard incompatible requests. (4) The paper's comparison includes a case where the sampling method does worse than the simple first-arrived heuristic (the clustered-demand instances), and it presents the learned policy as approaching, not reaching, an offline oracle that knows all future demands.