| paper | d4-conf-aaai-2013-0168 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper presents algorithms and an unnumbered correctness discussion, but no numbered theorem, lemma, corollary, or proposition stating a computational result. The proposed batch model is a statable extension, yet its repeated projects and cross-project capacity constraints add structure absent from the paper, so it cannot supply the missing anchor.
fails bit a — no named computational result to mirror
The proposed batch model covers cost-optimal plan selection across repeated projects; it leaves the paper's distributed VCG payment computation, privacy, and ex post faithfulness unmirrored.
There is no qualifying named computational result in the paper: the supplied text has Definitions 1 and 2, Algorithms 1–4, and an empirical table, but no numbered Theorem, Lemma, Corollary, or Proposition asserting a complexity result. The closest anchor is Definition 2, which specifies VCG for optimal MA planning: compute an optimal plan for \(\Pi\), optimal plans for each \(\Pi-\phi_i\), and payments from the resulting cost differences. The paper argues in unnumbered prose that Algorithm 4 is ex post faithful, and reports its performance, but neither gives an admissible anchor under your rule. So I cannot honestly present an anchor-backed case.
The strongest positive candidate I can offer is an unanchored mirror of the paper’s optimal-plan subproblem, rather than of its full VCG mechanism. Call it Continuous Batch Cost-Optimal MA-STRIPS. Take a finite set \(T\) of agent types, where a type fixes the agent’s complete MA-STRIPS action model, private state, public interface, and action costs. Give each type a rational supply \(\mu_t\), measured as agents available per project. The society consists of a unit mass of identical project instances, each with the same initial state, public facts, and goals. A configuration \(q\) is an ordinary legal MA-STRIPS plan for one project, with integral agent assignments; let \(a_t(q)\) count the type-\(t\) agents it uses and \(c(q)\) be its action cost. Choose a finite-support distribution \(x_q\geq 0\) minimizing \(\sum_q x_qc(q)\), subject to \(\sum_qx_q=1\) and \(\sum_qx_qa_t(q)\leq\mu_t\) for every \(t\). The output is the plan configurations and the fraction of projects assigned to each. Each project still receives a legal integral plan; only the type-level average can be fractional.
A plausible regime is a large programme of repeated infrastructure or facility projects. Many project instances share the same task template, while supplier teams fall into relatively few exact capability-and-cost types. Then the number of project instances is much larger than the number of types. This preserves the paper’s central computational object—finding cost-optimal joint plans from agents’ capabilities and costs—and makes type-level mass meaningful. A prover could attack it as a configuration LP whose pricing problem is to find a low-cost legal plan under type-resource prices.
I would expect general hardness to transfer from the planning structure: with ample type supplies, every project can use a legal plan for the original task, and averaging different plans cannot beat the cheapest one. Thus the continuous objective contains the underlying optimal-plan problem. That is a reason to expect a Class B result if the chosen MA-STRIPS task family is hard; the paper itself does not establish such a named complexity result. Further questions include whether the pricing problem is tractable for useful restrictions, and whether a continuum analogue of the paper’s leave-one-agent-out VCG payments can retain an individual incentive interpretation when one agent is an infinitesimal share of the population.
The weak point is substantial: the paper plans for a single shared task with finitely many agents, while this mirror adds a continuum of repeated tasks and capacity constraints across them. It covers cost-optimal plan computation, not the paper’s privacy-preserving distributed protocol or its strategyproofness claim. That is a plausible high-multiplicity extension of the authors’ problem, but without a qualifying named anchor it cannot satisfy the paper-specific anchor requirement.
The case against starts with the missing anchor. The paper defines the planning problem and VCG mechanism, gives Algorithm 4, proves protocol properties in an unnumbered correctness discussion, and reports benchmark timings. It states no named computational result about complexity or algorithms. Algorithm 4’s ex post faithfulness is the closest candidate, but that is a mechanism property, not a complexity result. The proponent’s proposed optimal-plan mirror is therefore an interesting extension, not a mirror anchored in a named result of this paper.
For a single MA-STRIPS task, the stronger obstacle is that agent multiplicity does not function as a resource. A type fixes the agent’s actions, state, and costs; the model puts no limit on how many actions one agent can perform. If several agents have the same type, one can perform the same action sequence as all of them. Once at least one agent of each type is present, increasing their masses does not change the legal plans or their costs. The problem depends on which types are present, not how much mass each has. Making the population atomless also removes the mechanism’s individual-pivotality question: deleting one agent changes no mass, so \(\Pi-\phi_i\) is unchanged. That cannot retain the paper’s leave-one-agent-out payments or their individual incentive interpretation.
The batch-plan proposal makes multiplicity matter by adding many projects and limiting how much agent supply can be assigned across them. But the original task has one shared plan and no such cross-project capacity constraint; its agents can perform multiple actions within that plan. Counting agents consumed by a configuration and constraining total supply therefore add a new resource-allocation problem. The resulting configuration LP may be worth studying, but its continuous object includes a mass of projects, and its central constraint comes from the new capacity model. It also drops the paper’s main mechanism question—how to compute payments and preserve individual incentives. That is a substantial neighboring problem, not a continuous version of Algorithm 4’s result.
The negative case has a limit: the paper’s discussion of competitive providers makes repeated projects and scarce teams imaginable, so a batch-planning model could be useful to these authors. I would not claim that no worthwhile high-multiplicity planning problem exists. The narrower case is stronger: on the paper’s single-task model, agent mass either collapses to type presence or erases individual pivotality; making mass consequential requires adding structure the paper does not study.
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.