{"id":"17f3f12a-fa00-4faa-9831-5c50b9255478","arxiv_id":"2501.18017","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A bilevel price-setting optimizer is integrated with Thompson sampling, allowing an energy community manager to learn individual prosumer signature weights from daily price responses without large historical datasets.","lead":"This paper embeds a bilevel price-setting model inside a Thompson sampling loop, so an energy community manager can learn how individual households respond to prices while setting daily prices. In a 25-prosumer simulation, the manager learns most household characteristics within five days and reaches near-zero regret within 100 days.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Misspecification of the signature set is the load-bearing risk: the numerical evidence is produced under the exact linear-mixture model the learner assumes, and Section 5 explicitly defers the incorrect-signature case.","rationale":"I read the paper in good faith. It develops a coherent integration of a bilevel price-setting MILP with Thompson sampling, and the numerical section is reproducible in principle. The central claim, however, is about learning real prosumer behavior. The strongest evidence is the regret curve in Figure 5, but that curve is produced by simulating true responses with the same signature models that the learner uses to update its posterior. This is the weakest point because the entire learning algorithm rests on Equation (2): the observed response is assumed to be a weighted sum of known signatures. If that assumption is wrong, the posterior update is not Bayesian for the true data-generating process, the learned weights are not individual characteristics, and the prices are optimized against a misspecified model. The paper explicitly lists 'specifying incorrect sets of signatures' as future work, which confirms the limitation is recognized but not addressed. The reader's weakest-assumption analysis identifies exactly this issue; I agree. I do not see a separate flaw that is more load-bearing: the lack of baselines and theoretical regret bounds are weaknesses but secondary, because even with baselines and bounds the self-consistent simulation would not establish robustness to signature misspecification. The proposed concrete test would settle whether the concern lands. Therefore the verdict remains CONDITIONAL, unchanged from the reader.","tokens_in":21723,"tokens_out":6386,"duration_ms":83567,"concrete_test":"Modify the Section 4.1 simulator so that the 'true' response of at least one prosumer is generated by an asset configuration not in the ten signatures, for example an EV with an availability window of 7:00-19:00, which lies between signatures 8 and 9. Run Algorithm 1 unchanged for 100 days, and compare cumulative regret and posterior predictive fit against the exact-model baseline run. If the final cumulative regret does not approach the near-zero level shown in Figure 5, or if the learned weights cannot predict the held-out responses, the central claim fails under realistic misspecification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a community manager can infer individual prosumer behavior without pre-existing data. In the simulation, the observation model (Algorithm 1, Step 4) generates the response as y = P θ* + ε, where P is calculated from the same signature lower-level problems (4) that the manager embeds in BiPS. The case study in Section 4.1 uses exactly these ten signatures to generate the 'true' prosumer responses. This makes the experiment a self-consistency check of the linear-mixture model, not a test of learning real prosumer behavior. Equation (2) and Section 2.2 assume the true response is exactly a weighted sum of a finite, known set of candidate signatures. A real household with an asset or habit not in the signature set, or with a nonlinear price response, violates this assumption. The posterior update (16)-(17) is Bayesian linear regression and is only valid under the assumed likelihood; under misspecification the posterior mean need not converge to anything meaningful, and prices optimized against the biased estimate need not achieve low regret. The paper states in Section 5 that 'specifying incorrect sets of signatures' is not investigated. Without a misspecification analysis, the 'individual-level learning' claim is conditional on the simulation's self-consistent assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework that combines bilevel price-setting optimization with online learning via Thompson sampling to learn individual prosumer behavior in an energy community. The manager solves a bilevel program (BiPS) each day, using sampled signature weights, then observes prosumer responses and updates Gaussian posterior distributions through Bayesian linear regression. The authors claim this is the first integration of bilevel programming into an online learning framework, with individual-level rather than aggregate-level learning. Numerical experiments with 25 prosumers and 10 signatures report near-zero cumulative regret after 100 days, fast learning of high-power assets within 5 days, and a prior-reset mechanism for non-stationary changes in signature weights.","tokens_in":21948,"tokens_out":2430,"duration_ms":29412,"significance":"If the central claim is supported, the paper offers a useful combination of two mature tools: KKT-based reformulation of lower-level LPs into a MILP, and Thompson sampling with conjugate Gaussian updates. The formulation is standard and the derivations are plausible: lower-level problems are linear, KKT conditions are sufficient, and the Fortuny-Amat reformulation is a recognized technique. The numerical case study is clearly described and the authors provide a code repository, which aids reproducibility. The main significance is the proposed integration itself and the demonstration that, under the exact linear-mixture model, the learner can recover signature weights and reduce regret quickly. However, the evidence for the practical claim of learning real prosumer behavior is limited by the self-consistent nature of the simulation, as detailed in the major comments.","major_comments":[{"comment":"The numerical evaluation is a self-consistency check rather than a test of learning under model misspecification. In Algorithm 1, Step 4, the 'true' response is generated as y = P θ* + ε, where P is computed from the same signature lower-level problems (Section 3.1.2) that BiPS embeds, and Section 4.1 uses exactly these ten signatures to generate the true prosumer responses. The paper's central claim that the community manager can infer individual prosumer behavior is therefore only demonstrated when the observation model coincides with the learner's model. Section 5 explicitly states that 'specifying incorrect sets of signatures' is not investigated. This is load-bearing: a real household with an asset or habit not represented by any candidate signature, or with a nonlinear price response, violates the linear-mixture assumption of Equation (2), and the posterior update (16)-(17) is not guaranteed to be meaningful under misspecification. The authors should add misspecification experiments (e.g., true prosumers whose assets are not in the signature set, or whose response is nonlinear in price) or an analysis of when the linear model remains adequate.","section":"Section 4.1 and Algorithm 1, Step 4"},{"comment":"The paper gives no theoretical regret bound. The central quantitative evidence is the empirical regret curve in Figure 5, but Section 3.3 explicitly says 'This paper does not discuss or analyze the theoretical bounds of regret under Thompson sampling.' Without any bound or even a stylized analysis, the claim that the algorithm achieves 'near-zero regret after 100 days' is an anecdotal observation from one simulation setup. At minimum, the authors should provide a regret analysis for a simplified version of the problem (e.g., a linear quadratic setting) or carefully qualify the claim as empirical and parameter-dependent.","section":"Section 3.3"},{"comment":"The MILP reformulation via the Fortuny-Amat approach is not fully specified. The paper states that the MPEC is reformulated using big-M, but it does not report the values of the big-M constants, the solver used, tolerances, or any validation that the complementarity conditions are satisfied at the reported solutions. Since the numerical results depend on the MILP being solved correctly, the omission of these details affects reproducibility. The authors should provide the big-M values or a systematic procedure for choosing them, and report solver settings.","section":"Section 3.1.3"},{"comment":"The identifiability of the signature weights is not discussed. Even under the exact linear-mixture model, observations y_n = P_n θ_n + ε may not uniquely determine θ_n if the columns of P_n are linearly dependent or if the price signals do not sufficiently excite the different signatures. The numerical study uses a specific set of signatures and prices, but no analysis is given for when the learning problem is well-posed. The authors should state conditions under which the posterior converges to the true weights, or acknowledge that the observed convergence may rely on particular features of the case study.","section":"Section 2.2 and Equation (2)"}],"minor_comments":[{"comment":"There is a typo in 'does not necessaneed to have perfectly learned' which should read 'does not need to have perfectly learned.'","section":"Section 4.4"},{"comment":"The caption says 'Each row of four subplots' and 'Each column of three subplots,' which is inconsistent with the actual layout; it should say three rows of four subplots or similar.","section":"Figure 6 caption"},{"comment":"The notation θ_k is used in the dual reformulation but the index k is not explicitly tied to prosumer n; the equation should clarify that these are the weights for the relevant prosumer, matching the definition in Equation (2).","section":"Equation (15)"},{"comment":"The reference list contains a typo 'Bejing' for Beijing in the Gopalan et al. entry.","section":"References"},{"comment":"The update equations (16)-(17) are written with Σ_{n,d} and R_n, but the covariance matrix of the response noise is introduced in Section 3.2 without an explicit definition of how R_n is chosen; a brief statement of the prior hyperparameters would improve reproducibility.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely within scope for Applied Energy given its application to demand response and energy communities. The novelty claim of being 'the first' to integrate bilevel programming with online learning is plausible but would benefit from a more careful comparison with the closest works (e.g., Tucker et al. 2020 and other Thompson-sampling pricing papers) to clarify the exact distinction. My main concern is that the empirical validation is entirely self-consistent with the modeling assumptions; the authors themselves defer the misspecification case to future work. This is fixable within the manuscript's scope by adding misspecification experiments and/or a theoretical consistency analysis, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper cleanly combines two mature ideas. The genuinely new bit is embedding a KKT-reformulated MILP of a leader-follower pricing problem inside a Thompson sampling loop, with per-prosumer signature weights. That combination is new relative to the cited bandit-pricing work, which uses simplified price-response functions. The math is sound: lower-level LPs are convex, KKT is necessary and sufficient, the Fortuny-Amat reformulation is standard, and the Bayesian linear regression updates are correct. Credit is due for shipping code and for saying plainly what the paper does not do, including the absence of regret bounds.\n\nThe main soft spot is exactly what the authors flag in Section 5: they never test what happens when the true prosumer behavior is not a weighted sum of the known signatures. The simulation generates data from the same linear mixture model the learner assumes, using the same ten signatures. That makes the experiment a self-consistency check, not evidence about real-world misspecification. This is not fatal, but it means the central claim—\"the community manager can infer individual prosumer behavior\"—is demonstrated only in the model's own terms.\n\nTwo smaller issues: there are no comparisons to simpler bandit baselines, so it is hard to tell how much of the regret performance comes from the bilevel structure versus the linear model and a well-chosen prior. And the big-M values are omitted, which is a reproducibility gap for the MILP.\n\nWho is this for: researchers working on demand response pricing or online learning for energy systems. A serious referee should engage; the gap is real and the method is plausible. The main request should be a misspecification experiment—generate true responses from a signature not in the set—and at least one baseline comparison. That would turn a self-consistency check into actual evidence.","headline":"A solid first integration of bilevel pricing with Thompson sampling, honestly scoped; the numerical evidence is a self-consistency check, so the practical claim rests on an unexamined misspecification assumption.","tokens_in":22489,"tokens_out":1555,"would_cite":true,"duration_ms":273289,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","91A65"],"pacs":[],"model":"deepseek-v4-flash","headline":"Dynamic pricing can learn each household's energy assets from observed price responses, eliminating the need for pre-existing datasets.","keywords":["dynamic pricing","demand response","online learning","energy communities","prosumer behavior","Thompson sampling","bilevel programming","capacity limitation"],"falsifier":"Run the algorithm on a simulated prosumer whose true asset is deliberately absent from the candidate signature set (for example, an EV with a fourth driving pattern) and check whether the posterior weights converge to any fixed combination and whether cumulative regret stays bounded; the paper's own Section 5 identifies this test as uninvestigated, and a result showing persistent regret would undermine the claim that individual characteristics are inferred.","tokens_in":21506,"feed_emoji":"⚡","tokens_out":5933,"duration_ms":51079,"temperature":0.7,"pith_summary":"This paper claims that dynamic pricing for demand response can work without assuming the grid operator already knows how each household uses its energy assets. It combines bilevel price optimization with Thompson sampling, so the community manager sets prices each day and learns the 'signature weights' describing each prosumer's assets and habits from their observed responses. The motivation is concrete: an energy community that must cap its grid import to earn a discounted tariff can learn enough about its 25 prosumers to deliver that service after roughly a week, and reach near-zero regret within 100 days. If the claim holds, online learning removes a major practical obstacle to price-based demand response, since no large pre-existing dataset on individual behavior is required.","feed_headline":"A price framework that learns each home's energy assets in days","feed_subtitle":"Community managers can cut grid peaks without pre-existing data on household behavior.","key_machinery":"The load-bearing mechanism is the linear signature-response model $y_n = P_n \\theta_n + \\epsilon$: a prosumer's observed daily profile is a weighted sum of $K$ pre-specified power signatures, and the unknown weights $\\theta_n$ are what the manager must infer. On that model sits a Thompson-sampling loop in which Gaussian priors over $\\theta_n$ are sampled daily, the samples feed the bilevel price-setting problem BiPS, and Bayesian linear regression updates (16)-(17) refine the posterior after each day's metered responses. The bilevel problem itself is turned into a single-level MILP by replacing the prosumers' lower-level linear programs with their KKT conditions and using strong duality to linearize the bilinear price-response terms, so the whole price-setting step is solvable at each iteration.","core_discovery":"The paper's central claim is that this is the first framework to embed bilevel programming inside an online learning loop, enabling individual-level learning of prosumer behavior rather than aggregate-level learning. Each prosumer's daily load is modeled as a weighted sum of a finite set of candidate 'signatures' (flexible baseload, PV, battery, heat pump, and EV with specified usage patterns), with unknown weights; the manager maintains a Gaussian posterior over these weights, samples them each day, solves a bilevel price-setting problem (reformulated as the mixed-integer linear program BiPS via KKT conditions, strong duality, and the Fortuny-Amat big-M reformulation), and updates the posterior by Bayesian linear regression after observing true metered responses. In a 25-prosumer simulation with 10 signatures per prosumer over one year of Danish price, weather, and PV data, the algorithm achieves near-zero cumulative regret after about 100 days, learns the high-power asset signatures (EVs and batteries) within about five days, and restores learning after a non-stationary change in weights if the prior is reset.","pith_inferences":["If the claim holds, the same learning loop could be applied beyond capacity limitation to any flexibility service where a leader sets prices and observes followers' responses, a generalizability the paper itself notes.","The strongest practical risk the paper leaves open is misspecification of the signature set: if a household's real assets cannot be expressed as a weighted sum of the candidate signatures, the learned weights need not reflect reality, and the numerical guarantees would not carry over.","Because the learned weights encode occupancy and driving habits, the framework turns smart-meter data into inferred private information, making privacy-preserving variants like differential privacy or federated learning a natural next step.","The paper provides numerical evidence but not theoretical regret bounds; a proof of Bayesian regret for this bilevel Thompson-sampling setup would strengthen the claim of general applicability."],"forward_implications":["A community manager can offer capacity-limitation services to a DSO without asking households for private usage data; observing responses to prices suffices.","High-power assets (EVs and batteries) are identified within about five days, removing the largest early risk of exceeding the grid capacity limit.","Near-zero cumulative regret after 100 days means the learned prices eventually match the prices a clairvoyant manager with full knowledge would set.","If a prosumer's actual behavior changes, resetting the prior distribution restores learning within roughly the same time frame as the original learning.","Individual-level price setting follows from individual-level learning, enabling more tailored demand response than aggregate-level pricing."],"supporting_citations":[{"why":"Supplies the Thompson sampling tutorial and algorithmic foundation that the integrated learning loop is built on.","marker":"Russo et al., 2020"},{"why":"Provides the generalized battery model used to formulate the heat pump signature's lower-level problem.","marker":"Hao et al., 2015"},{"why":"Extends constrained Thompson sampling to real-time electricity pricing, the closest prior setting the paper builds on.","marker":"Tucker et al., 2020"},{"why":"Cited as the source for theoretical regret bounds that justify using Thompson sampling in complex online problems.","marker":"Gopalan et al., 2014"},{"why":"Supplies the flexible baseload demand profiles used as inputs in the numerical case study.","marker":"McKenna et al., 2020"},{"why":"Supplies the methodology for generating the PV production data used as an exogenous input in the simulations.","marker":"Pfenninger and Staffell, 2016"}],"fun_headline_variants":["Learn each home's energy assets in days without prior data","AI pricing that learns household energy habits on the fly","No historical data needed: prices reveal prosumer behavior in days","Bilevel pricing meets online learning: homes learned in 5 days"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole learning loop assumes that each prosumer's true response is exactly a weighted sum of the finite set of candidate signatures; if a household's actual assets or habits are not captured by any of these signatures, the learned weights and the prices derived from them need not be correct.","fun_headline_variants_meta":{"raw":{"variants":["Learn each home's energy assets in days without prior data","AI pricing that learns household energy habits on the fly","No historical data needed: prices reveal prosumer behavior in days","Bilevel pricing meets online learning: homes learned in 5 days"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000868,"raw_usage":{"total_tokens":3743,"prompt_tokens":908,"completion_tokens":2835,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":2765}},"tokens_in":524,"tokens_out":2835,"duration_ms":21685,"temperature":1.0,"reasoning_tokens":2765,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:58:56.702340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on a simulated prosumer whose true asset is deliberately absent from the candidate signature set (for example, an EV with a fourth driving pattern) and check whether the posterior weights converge to any fixed combination and whether cumulative regret stays bounded; the paper's own Section 5 identifies this test as uninvestigated, and a result showing persistent regret would undermine the claim that individual characteristics are inferred.","supporting_citations":[{"cited_title":", author Sanandaji, B.M","cited_arxiv_id":null,"evidence_quote":"Provides the generalized battery model used to formulate the heat pump signature's lower-level problem."},{"cited_title":", author Moradipari, A","cited_arxiv_id":null,"evidence_quote":"Extends constrained Thompson sampling to real-time electricity pricing, the closest prior setting the paper builds on."},{"cited_title":", author Mannor, S","cited_arxiv_id":null,"evidence_quote":"Cited as the source for theoretical regret bounds that justify using Thompson sampling in complex online problems."},{"cited_title":", author Thomson, M","cited_arxiv_id":null,"evidence_quote":"Supplies the flexible baseload demand profiles used as inputs in the numerical case study."},{"cited_title":", author Staffell, I","cited_arxiv_id":null,"evidence_quote":"Supplies the methodology for generating the PV production data used as an exogenous input in the simulations."}],"review_version":1}