| paper | d4-conf-ijcai-2021-0014 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
Theorem 5.2 gives a closed-form guarantee, but it does not state a computational result under the required anchor rule. The proposed population version is recognizable, yet computing \(q_t\) from supplied per-type inputs is just \(O(\tau)\) arithmetic, while deriving continuum contribution values would require a new model.
fails bit a — no named computational result to mirror
The candidate mirror covers Theorem 5.2 and Section 5’s redistribution formula; it leaves the dispatch ILP, fairness objectives, and empirical results alone.
The strongest positive case is an end-of-day income-pooling problem for a large city fleet. There is a qualification, though: the paper has no numbered result classifying a computational problem as tractable, hard, FPT, or similar. The matching ILP in Section 3.2 is used in the experiments, but it is not a numbered result. The closest named anchor is Theorem 5.2, proved here: holding a driver’s value \(v_i\), risk parameter \(r\), and total pre-redistribution income fixed, the minimum possible post-redistribution income \(q_i\) is \(\min(r v_i,(1-r)v_i)\). That is a wage-floor guarantee, not a complexity result. So it does not meet the requested anchor rule literally.
As a substantive candidate, I would mirror Theorem 5.2 in a fleet with many drivers and recurring operational states. Let a type \(t\) specify the driver attributes relevant to matching and redistribution—such as vehicle capacity, location, current route state, and work history. Let \(\mu_t\) be the fraction of the fleet of that type, and let \(p_t\) and \(v_t\) be its per-driver earnings before redistribution and per-driver contribution value. Drivers with the same type must be indistinguishable in all attributes the problem uses, including \(p_t\) and \(v_t\). The plausible regime is a large fleet whose drivers occupy relatively few repeated states on the platform’s discretized service network; it is not every ride-pooling market.
To carry the paper’s Shapley idea into this setting, \(v_t\) could be a type’s marginal value per unit of driver mass in a continuum coalition game. For example, if \(\Pi(z)\) is the income produced by a coalition with mass profile \(z\), an Aumann–Shapley path value would assign
\[ v_t=\int_0^1 \frac{\partial \Pi(\alpha\mu)}{\partial z_t}\,d\alpha. \]
Assume \(\sum_t\mu_t p_t=\sum_t\mu_t v_t\). For \(r<1\), define post-redistribution per-driver income by
\[ q_t = r v_t+ \frac{\max(0,v_t-rp_t)} {\sum_s\mu_s\max(0,v_s-rp_s)} (1-r)\sum_s\mu_s v_s. \]
The continuous problem is: given a finite-support rational distribution \(\mu\), the per-type inputs \(p_t,v_t\), and \(r\), compute every \(q_t\) and determine whether every supported type satisfies \(q_t\geq\min(r v_t,(1-r)v_t)\). A solution is the vector of type-level incomes together with the bound for each type. At \(r=1\), set \(q_t=v_t\).
This is a close mirror of the paper’s redistribution step: the population is the drivers, mass is fleet share, and the action is reallocating the pooled income across driver types. With \(p_t\) and \(v_t\) supplied, computing \(q_t\) is tractable in \(O(\tau)\) arithmetic operations, and the wage-floor guarantee appears to extend by replacing finite sums with mass-weighted sums. That is a modest tractability result, though: the difficult question may be obtaining the \(v_t\) values from the ride-pooling game, not carrying out the redistribution once they are known. This covers Theorem 5.2 and Section 5’s redistribution model; it does not cover the dispatch ILP, the fairness objectives, or the empirical findings.
My weakest point is the continuum Shapley value. The paper defines driver value through marginal contributions across subsets of named drivers; an Aumann–Shapley value is a plausible mass-based analogue, but the paper neither defines it nor proves it is the limit of its finite Shapley values. Also, if earnings and contribution values are sufficiently individualized, the number of distinct types may approach the number of drivers. That makes this a credible candidate mirror, but not a strong computational-complexity anchor under the programme’s stated test.
The strongest objection is that the paper has no named computational result to mirror. Its matching ILP and approximate dynamic program are methods used in experiments, not complexity results; Theorems 5.1 and 5.2 are claims about income redistribution. That leaves no result whose tractability, hardness, or approximation landscape continuization could chart.
The proposed Theorem 5.2 mirror does not supply that missing target. Given the per-type inputs \(p_t\) and \(v_t\), its \(q_t\) values require only a weighted sum and one calculation per supported type: \(O(\tau)\) arithmetic operations. The wage floor follows immediately from the formula: its redistribution term is nonnegative, so \(q_t\ge r v_t\ge\min(r v_t,(1-r)v_t)\), for nonnegative \(v_t\). Replacing the finite sum over drivers with a mass-weighted sum preserves the formula, but adds no computational question of substance.
The strongest repair is to ask how to compute the values \(v_t\). But the paper defines them through Shapley values over subsets of named drivers, with each coalition’s income obtained by rerunning the matching problem. An Aumann–Shapley path integral requires a new fractional-coalition game, a choice of how requests scale with fleet mass, and a reason that this value is the right limit of the paper’s Shapley value. Those choices could define a worthwhile problem, but they are not consequences of Theorem 5.2; they supply a new model and the hard question the paper itself did not pose.
A fluid dispatch model is another plausible repair: distribute drivers over route and location states, represent requests by origin, destination, and time, and optimize flows over those states. Yet the paper’s driver state includes current routes and income history, and its rider requests arrive over time and compete across feasible matchings. An exact type description can therefore have support that grows with the fleet and horizon. A useful low-support version would need additional assumptions about repeated states and demand, plus a reformulation of the dynamic objective. That may be a promising project, but it is not a mirror anchored in a result of this paper.
So the negative case is strong for the proposed anchor and for treating this paper as a direct ChoCo complexity target: its only named theorem yields a straightforward mass-weighted restatement, while a nontrivial computational question requires building new machinery. The universal claim is weaker. A carefully specified large-fleet fluid dispatch problem could be worthwhile; the paper alone does not establish that it would be.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.