Freight carriers that share transfer hubs can save a lot by cooperating: they pool loads, cut empty miles, and swap backhauls. The models that estimate these savings, and the bargaining rules that divide them, almost always assume that hubs and fleets have unlimited capacity. That assumption is not just a numerical approximation. It also decides how much each carrier could earn on its own, and that "go-it-alone" payoff is the baseline every fair-division rule is built on. If hubs are actually finite, carriers acting alone would crowd each other out. Their real outside options are then worse than the models claim. The surplus credited to cooperation, and each carrier's share of it, are both mismeasured, and a deal that looked worthwhile on paper can leave some carrier worse off once trucks queue at a full yard.
We propose to re-derive cooperative surplus sharing among independent carriers with the hub capacity limit put back in. The key observation is simple. If we attach a price to each hub's capacity in each time period, every carrier's planning problem splits back into the fast, independent problem it already solves under the infinite-capacity assumption, now with tolls added. The infinite-capacity model is exactly the special case where all hub prices are zero. The same prices then define a disagreement point (what each carrier would get if talks collapse) that is consistent with finite capacity, and the bargain is computed on that basis. Each carrier keeps its own solver and reveals only how much hub capacity it wants at a given price.
Core Claim
The infinite capacity assumption distorts cooperative allocation through the disagreement point, not only through the plan. Pricing the shared capacity through dual variables corrects both, and it keeps the decoupling that made the assumption attractive. Concretely, let carrier $i$ choose a plan $x_i$ with profit $\pi_i(x_i)$. Let the plan occupy $a_{ht}(x_i)$ staging slots at hub $h$ in period $t$, against capacity $C_{ht}$. Let $\tau_i$ be a monetary transfer to carrier $i$ and $d_i$ its disagreement payoff. The capacity-aware bargain is
$$\max_{x,\tau} \sum_i \log\big(\pi_i(x_i) + \tau_i - d_i\big) \quad \text{s.t.} \quad \sum_i a_{ht}(x_i) \le C_{ht}\ \ \forall h,t, \qquad \sum_i \tau_i = 0.$$
Relaxing the capacity constraints with prices $\lambda_{ht} \ge 0$ leaves each carrier with $\max_{x_i} \pi_i(x_i) - \sum_{h,t}\lambda_{ht}\, a_{ht}(x_i)$. That is its existing infinite-capacity problem with hub tolls, and it is the familiar problem when $\lambda = 0$.
The claim has a falsifiable empirical consequence. Suppose bargains are struck using infinite-capacity disagreement points and then executed under real hub limits. We predict they will violate individual rationality for a measurable fraction of carriers, meaning some carrier ends up below its go-it-alone payoff. We also predict this fraction grows with hub utilization. The capacity-consistent version should eliminate those violations at little cost to total profit.
There is a point we must be honest about from the start. When side payments are allowed, the Nash bargaining solution reduces to two steps: maximize total profit, then give each carrier its disagreement payoff plus an equal share of the gain. So the content of this proposal lives in two places. The first is how $d_i$ is computed. The second is whether total profit can be maximized under capacity limits without a central planner. It does not lie in the product form itself. We also study the case without side payments, where carriers trade hub slots but not money. There the product form changes the allocation.
Relation to Prior Work
The proposal sits on a gap that several papers mark from different sides.
Three papers motivate the problem by relying on the assumption and naming it as the weak point. Scalability and Performance of Decentralized Planning in Flexible Transport Networks shows that decentralized dispatch agents trained by approximate dynamic programming scale where a central planner fails. It explicitly assumes infinite hub capacity and includes no surplus sharing. Optimizing Autonomous Transfer Hub Networks: Quantifying the potential impact of self-driving trucks solves a national single-fleet planning model over 100 hubs and leaves hub staging limits as future work. Non-Cooperative Game of Coordinated Scheduling of Parallel Machine Production and Transportation in Shared Manufacturing finds selfish equilibria and lists limited vehicle availability as unmodeled.
The closest work is Incentive design and profit sharing in multi-modal transportation networks. It already splits multi-operator transport profit with an asymmetric Nash bargaining solution, setting each $d_i$ to pre-cooperation profit. However, its passengers move as continuous flows with no vehicle or node caps. We differ in exactly the place that matters: our disagreement points and our cooperative plan both respect hard, discrete hub capacity, and we measure what goes wrong when they do not.
The method comes from Resource Sharing in the Edge: A Distributed Bargaining-Theoretic Approach.. That paper takes the logarithm of the Nash product and solves it by exchanging prices between edge computing providers, without revealing internal workloads. We reuse that machinery. Its setting, though, is convex and divisible server capacity. Our per-carrier problems are integer routing and scheduling problems, so a duality gap appears, and handling it is part of the work.
Ridesharing and Fleet Sizing For On-Demand Multimodal Transit Systems constrains our design from the other side. It shows that putting finite capacity back into transport design is tractable centrally, but only after careful reformulation and under a single owner. That makes a centralized solve with full information the natural reference for our decomposition.
Research Program
The work has four parts.
Shared-hub benchmark. We will build a multi-carrier benchmark with finite hubs. The order and network structure come from the synthetic multi-carrier benchmark of van Heeswijk and La Poutré, which has 3, 5, 10 and 25 agents and up to 3 arcs per node. We add per-period staging capacities $C_{ht}$, set at several fractions of the peak load observed under infinite capacity. Where the Ryder full-truckload data used by Lee et al. can be obtained, we will partition its 6,842 orders over 100 hubs among synthetic carriers. If it cannot be obtained, we will say so and generate orders with matching structure.
Per-carrier solver. Each carrier's subproblem is its own toll-adjusted plan. Small instances are solved exactly by integer programming. Large instances use a learned dispatch policy that takes the prices as an additional input.
Price coordination and repair. A coordinator, or the carriers among themselves, updates hub prices from observed overuse using projected subgradient steps. Because the subproblems are integer, the prices alone may not clear capacity. A slot-repair step reassigns the remaining conflicts. We report the gap between the relaxed and the repaired plan.
Capacity-consistent disagreement point. For each carrier, $d_i$ is its payoff in the non-cooperative outcome: all carriers plan selfishly and hub slots are rationed first-come-first-served. This is computed by iterated best response. We contrast it with the infinite-capacity $d_i$, which is each carrier's solo optimum with $\lambda=0$. For the setting without side payments, transfers are removed and carriers bargain directly over slot allocations.
Analysis. We will characterize when the surplus attributed to cooperation rises or falls once capacity binds. Two effects pull in opposite directions. Cooperation shrinks because the joint plan has to fit, and outside options shrink because carriers congest each other alone. The net effect is the empirical question.
Baselines
Naive reference. This is today's default practice. Each carrier plans alone under infinite capacity, hubs ration conflicts first-come-first-served, and there is no cooperation. It shows the cost of ignoring both capacity and cooperation.
Infinite-capacity bargain, executed under capacity. This is the strongest existing method on the allocation question. We will use the bargaining protocol of Incentive design and profit sharing in multi-modal transportation networks, transplanted from passenger flows to our freight benchmark. The joint plan and the $d_i$ are computed with $\lambda=0$, and the plan is then forced through the real hub limits by first-come-first-served repair. That paper reports a 74% platform profit increase from cooperation in one simulation, but under uncapacitated flows; no reported result found on Lacuna for it under hard capacity. Fairness of comparison is straightforward: it sees the same orders, uses the same per-carrier solver, and gets the same compute budget. We beat it if we produce fewer individual-rationality violations at equal or higher realized total profit.
Centralized optimum. This is the strongest method on the efficiency question. A single solver sees every carrier's private data and directly maximizes the log Nash product under capacity, in the spirit of the centralized reformulation in Ridesharing and Fleet Sizing For On-Demand Multimodal Transit Systems. It is run with the same wall-clock limit, 24 hours as in van Heeswijk and La Poutré. On their benchmark, decentralized planning came within 3% of the centralized planner at 3 to 5 agents, the central planner did not finish one iteration in 24 hours at 10 and 25 agents, and the learned policies beat shortest-path and just-in-time heuristics by 20% to 26%. All of those numbers are under infinite hub capacity. Matching that 3% at small sizes is our bar, and finishing at 25 agents is our claim.
Ablation. This baseline uses capacity prices for the plan but infinite-capacity disagreement points. It isolates the proposal's key ingredient, which is the disagreement point.
Why not just…?
Why not just solve one centralized model with hard capacity? On small instances it is the gold standard, and we use it as one. Its weaknesses are practical. Every carrier would have to disclose costs and orders, which the edge-sharing and multi-modal bargaining work treats as a deal-breaker. It also scales poorly: van Heeswijk and La Poutré report the central planner stalling at 10 agents. Our experiments will report the optimality gap between our decomposition and the centralized model. If that gap is large at small sizes, the decomposition is not worth its privacy benefit, and we will say so.
Why not just charge congestion tolls at hubs and skip bargaining? This is a real threat, because tolls alone may already deliver an efficient plan. What tolls leave unanswered is who receives the toll revenue on a hub that no single carrier owns, and whether every carrier ends up above its outside option. We test this directly. One condition uses tolls with revenue returned in equal shares, and we compare its individual-rationality violations against the full bargain. If equal rebates already satisfy everyone across utilization levels, the bargaining layer adds nothing, and the contribution narrows to the disagreement-point analysis.
Why not just split gains by Shapley value or in proportion to volume? The Shapley value requires the optimal capacity-constrained plan for every one of the $2^n$ possible coalitions. We will compute it at 3 and 5 carriers to compare allocations, but it cannot be computed at 25. Proportional splits are cheap, but they ignore outside options entirely. Our individual-rationality metric will show whether that matters in practice.
Proposed Experiments
Metrics. We track four quantities across hub utilization levels, from slack to tight. Realized total profit is measured after capacity repair. The individual-rationality violation rate is the share of carriers whose realized payoff falls below their capacity-consistent outside option. Fairness is measured by Jain's index, for which Resource Sharing in the Edge: A Distributed Bargaining-Theoretic Approach. reports values above 0.95 in its convex setting. We also record runtime and messages exchanged.
First decisive check. At tight capacity, the infinite-capacity bargain should show nonzero and rising individual-rationality violations, while the capacity-aware bargain shows near zero. If violations stay near zero even for the infinite-capacity bargain, the core claim is wrong.
Second decisive check. The ablation should account for most of the gain. If pricing the plan alone fixes the violations, the disagreement-point argument fails.
Third decisive check. At 3 to 5 carriers, the decomposition with slot repair should come within a few percent of the centralized optimum. We also report the duality gap introduced by integer subproblems. If the repair gap exceeds the cooperative gains, the method fails in this regime.
References
- Scalability and Performance of Decentralized Planning in Flexible Transport Networks
- Optimizing Autonomous Transfer Hub Networks: Quantifying the potential impact of self-driving trucks
- Non-Cooperative Game of Coordinated Scheduling of Parallel Machine Production and Transportation in Shared Manufacturing
- Incentive design and profit sharing in multi-modal transportation networks
- Resource Sharing in the Edge: A Distributed Bargaining-Theoretic Approach
- Ridesharing and fleet sizing for On-Demand Multimodal Transit Systems