| paper | d4-conf-atal-2019-0191 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | unclear |
Theorem 1 gives a POMDP representation, but the paper neither uses that representation to compute its policy nor makes its method depend on solving it. The experiments and numbered observations do not meet the required named computational-result gate, so the proposed fleet problem cannot supply the missing anchor.
fails bit a — no named computational result to mirror
The proposed fleet model extends Theorem 1's EE-to-POMDP representation to households with different types; it does not mirror a computational result established in the paper. The GP/EVOI experiments and HVAC performance comparisons are left without a continuous-population counterpart.
There is a plausible population mirror here, but the paper gives me no anchor that clearly satisfies your rule. The closest is Theorem 1, proved in this paper (with its full proof in the supplement): every experiential elicitation instance can be formulated as a POMDP. Its construction puts the unknown reward function into the hidden state, adds queries to the action set, leaves the embedded reward fixed under transitions, and makes query observations follow the response distribution. The POMDP belief state is then the posterior over possible reward functions. The paper discusses why exact POMDP solution is often intractable, but Theorem 1 itself proves a representation, not a complexity classification; and the experiments do not compute by solving that POMDP. So, under the stated anchor gate, it is not a qualifying computational anchor. The numbered Observations do not supply one either.
If Theorem 1 is accepted as a looser formulation anchor, I would propose Fleet Experiential Elicitation\(_\infty\). Consider a large HVAC fleet with a finite set \(T\) of complete household types: a type fixes the resident’s reward function, query costs, and response model. The instance gives rational masses \(\mu_t\) summing to \(1\), with \(\tau=|\operatorname{supp}(\mu)|\) small compared with the number of homes. Residents’ types are hidden, but their HVAC states and query histories are observable. A controller can query a fraction of homes sharing an observable history, or choose HVAC actions for that cohort. Query responses split a cohort according to the type-dependent response distributions; HVAC states evolve according to the known transition model.
The underlying problem is: given finite state, action, and query sets, the transition model, type-specific rewards and query costs and response distributions, the masses \(\mu_t\), and a threshold \(B\), does there exist a query-and-control policy whose expected discounted aggregate resident reward minus query cost is at least \(B\)? A solution is such a policy, mapping observable cohorts and their histories to queries or control actions. This keeps the paper’s central coupling between learning preferences and controlling the system. It changes a single home into a fleet, but multi-building HVAC management gives that change a credible setting: many homes can share preference and response profiles while remaining distinct in their current states and experiences.
For this candidate, I would tentatively expect exact optimization to be hard, leaning toward Class C: hidden type composition and response-driven cohort splitting can make the policy problem complex even when \(\tau\) is moderate. That is a forecast, not something established by this paper. The main questions it opens are whether the aggregate belief has a tractable finite-dimensional representation, whether a finite-fleet version converges to this mass model, and whether any hardness survives when the number of types is fixed.
The soft spot is substantial: Theorem 1 does not classify the complexity of the problem, and the fleet model replaces learning one user’s reward with learning how to serve a distribution of residents. The paper’s authors might recognize the shared experiential elicitation and control problem, but this is not a clean mirror of a named hardness or tractability result in their paper. Thus the positive case is a credible research direction, but it does not meet the requested anchor standard as written.
There is no computational result in this paper for a continuous population to mirror. Theorem 1 shows that an experiential elicitation instance can be represented as a POMDP; it does not classify the complexity of optimal elicitation or prove an algorithmic bound. The numbered observations give reductions or policy facts, and the experiments compare heuristics. Even if Theorem 1 is accepted as a loose anchor, changing one user into a population does not continue its result: the theorem’s representation already allows uncertain rewards, while it says nothing about the new fleet problem’s complexity.
The proposed fleet problem has a sensible starting point: many homes can share preference and response profiles. But if homes act independently, have independent priors and feedback, and face no shared constraint, their aggregate objective is just a sum of individual EE objectives. The population mass scales that sum; it adds no new computational question. To make the fleet problem interesting, the controller needs some cross-home coupling, such as pooled information, a query budget, or a shared grid constraint. Those choices define a new multi-user control problem rather than a population version of a result in this paper.
The strongest fleet model still has a hard modeling fork. Residents’ query costs and response noise depend on their distinct state histories. Aggregate by fixed preference and device types, and those histories can produce different query responses and costs within each type. Track them exactly, and query histories—including real-valued responses—can generate a growing, potentially continuous distribution of information states. The model then needs machinery for controlling that distribution. Alternatively, require identical experiences and shared observations, and it largely collapses to a representative user. A continuum can also make a single query’s normalized cost vanish while preserving its informational value; requiring queries to a positive mass avoids that particular degeneracy but changes the feedback being modeled. Each choice is defensible, but it is a substantive new specification, not a mirror supplied by the paper.
So the negative case is strong against the requested anchor standard: none of the paper’s named results establishes a computational classification to continuize, and the proposed fleet version either scales independent instances or adds new population-level structure. It is weaker as a universal claim. A fleet problem with shared query budgets or pooled learning could be worthwhile on its own; I cannot honestly rule out every such scenario. I would describe it as a follow-on research direction, not a continuous mirror of this paper’s computational results.
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.