REVIEW 3 major objections 5 minor 39 references
Generalising Battery Control in Net-Zero Buildings via Personalised Federated RL
T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Federated TRPO without tuning is claimed to match tuned PPO for net-zero battery control.
desk verdict The headline claim contradicts the paper's own tables; the useful toy benchmark and honest experiments still merit a serious referee. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The mechanism carrying the argument is a federated learning loop in which each building's agent samples its own trajectories and the gradients are averaged across agents, which the authors equate to FedAvg aggregation, plus a split-learning personalization block (an optional private encoding of household-specific features) that is not shared. TRPO's constrained, second-order policy updates provide stable, adaptive step sizes without manual tuning, while PPO's clipped surrogate serves as the tuned baseline. The environment is a stripped-down CityLearn scenario with synthetic data engineered so that solar generation plus battery storage can cover each home's load completely, making the optimal policy known (zero cost and zero emissions).
What would settle it
Run the same experiment in a true federated setting: each building trains its own model on its own local data for several local epochs, then sends only model weights to a central server for averaging, with no direct sharing of observations across buildings, and compare the evaluation reward against the paper's stacked-gradient results; if true federated TRPO does not match a tuned PPO, the central claim is refuted.
Extended reading notes
Core claim
The paper's central claim is that Federated TRPO, combined with FedAvg-style aggregation and a personal encoding block kept private to each building, is comparable with state-of-the-art federated RL methodologies without hyperparameter tuning. In the authors' experiments, a tuned PPO generally achieves higher rewards and lower cost and emissions, but TRPO converges reliably without tuning and consistently beats the no-battery baseline in both two-building and five-building microgrid scenarios. The paper further reports that grouping features and personal encoding help TRPO in the shifted (more heterogeneous) datasets. The authors conclude that even a simple optimal policy—charge when solar is available, discharge when needed—is hard to learn, and they suggest a pragmatic warm-start strategy: a few PPO iterations followed by TRPO.
Load-bearing premise
The paper assumes that training a single network on observations stacked from all buildings is equivalent to doing actual federated averaging of locally trained models, and the entire 'federated' claim falls if that equivalence does not hold in practice.
Editorial extensions
If this is right
- If the paper is right, an untuned TRPO can be a drop-in replacement for tuned PPO in federated building-control applications, saving the effort of hyperparameter search.
- The demonstrated generalization across buildings suggests that privacy-preserving federated learning can pool knowledge across heterogeneous sites and still approach net-zero operation.
- The finding that simple optimal policies are hard to learn supports curriculum-style training that starts with well-controlled toy scenarios before scaling to real-world complexity.
- Warm-starting with a few PPO iterations and then switching to TRPO could yield a practical hybrid algorithm for battery control.
- The synthetic-data testbed with a known optimal policy offers a diagnostic for measuring optimality gaps in RL algorithms beyond this specific application.
Reading between the lines
- A reader should note that the equivalence between the paper's stacked-observation setup and a true federated protocol (with local training steps and communication rounds) is assumed, not tested, so the 'federated' results may not transfer directly to real federated deployments with communication constraints or privacy noise.
- The reported tables show a large reward gap between tuned PPO and untuned TRPO; the claim of parity may depend on the definition of 'comparable,' and a fair test would tune TRPO as well or hold both methods to the same tuning budget.
- The synthetic scenario's design—where net-zero is achievable but difficult to learn—could be reused as a benchmark for diagnosing convergence and optimality of other RL algorithms, especially second-order methods.
- If the stacked-gradient approximation is found to be inaccurate, the paper's conclusions about federated TRPO would need to be re-verified under a real federated averaging scheme, which is a straightforward extension.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies battery control in net-zero building microgrids using a customized CityLearn environment with synthetically generated data and a known optimal policy. It compares PPO and TRPO under several collaborative configurations (shallow network, personal encoding, grouped features, and shifted data) and reports that an untuned Federated TRPO is comparable with a tuned federated PPO. The authors also contribute a TorchRL wrapper for CityLearn and release code.
Significance. If the central claim were supported, the paper would offer a useful practical lesson: a hyperparameter-free TRPO could match a tuned PPO in a privacy-preserving shared-control setting for building energy management. The paper has concrete strengths: a reproducible synthetic benchmark with a known optimal policy, five-seed experiments, public code, and a modular environment wrapper. However, the headline claim is directly contradicted by the paper's own quantitative tables, and the experiments do not actually implement a federated protocol. The significance of the work as a federated RL contribution is therefore not established.
major comments (3)
- [Abstract, Section 4.2, Tables 1 and 2, Section 5] The central claim that Federated TRPO is comparable with tuned federated PPO is contradicted by the paper's own results. In Table 1, the best TRPO evaluation reward is -0.47064 (pe) while the best PPO evaluation reward is -0.14021 (pe gf); in Table 2, the corresponding values are -0.45768 (base) for TRPO and -0.14662 (pe gf) for PPO. Section 4.2 explicitly states that "the performance of the tuned PPO was superior to that of the TRPO." No standard deviations or statistical tests are reported for these five-seed averages, so the conclusion "Federated TRPO performed on par with a tuned federated PPO" has no statistical or numerical basis in the presented data.
- [Appendix A.2] The paper argues that "passing the stacked observations to a single network has the same effect as doing a FedAvg," but this equivalence is not established and is generally false. FedAvg involves multiple local updates between communication rounds, client sampling, and aggregation of model weights; a single forward/backward pass over a concatenated batch of agent observations is a centralized mini-batch update, not a federated protocol with local training steps. Since no communication rounds, partial participation, or privacy mechanism are implemented or analyzed, the claims that the framework is "privacy-preserving" and that the results concern "Federated TRPO" are not supported by the experiments.
- [Abstract and Section 5] The abstract claims comparability with "state-of-the-art federated RL methodologies," but the paper contains no comparison with any external federated RL baseline. All comparisons are among PPO and TRPO variants within the proposed setup. The conclusion's additional recommendation to warm-start with PPO and switch to TRPO is also not tested anywhere in the experiments.
minor comments (5)
- [Tables 1 and 2] The tables report means over five seeds but no standard deviations or confidence intervals, even though the accompanying figures show substantial variance; please add dispersion measures to the tables.
- [Table 2 caption] The caption reads "Values in the table represent the average for 5" and appears to be missing the word "seeds."
- [Section 4.2] The sentence "PPO overperforms TRPO significantly, but we highlight that we didn't spend time fine-tuning it" is ambiguous about whether "it" refers to TRPO or PPO; Appendix C indicates PPO was tuned while TRPO was not, so please rephrase.
- [Appendix B] The description of the selling-price margin as "a margin of 40% of the lowest daily electricity pricing" is unclear; specify exactly how the selling price is computed from the daily pricing series.
- [Section 3.2 and Appendix A.2] Figure 1 labels the Personal Encoding block as a local component, but Appendix A.2 states that "we are not in the process of training a private part of the policy network" and uses a one-hot encoder. Please clarify what the personal encoding block actually contains and how it relates to split learning.
Circularity Check
No derivation-chain circularity; the 'Federated' claim is self-definitional because Appendix A.2 equates centralized stacked training with FedAvg.
-
self definitional
[Appendix A.2 (Federated RL)]
"Passing the stacked observations to a single network has the same effect as doing a FedAvg: each agent samples their trajectories, and after a forward pass, the backward pass will propagate the average gradient of all of them, which is, in practice, the same behavior we expect from the FL methodology."
The advertised central result is about 'Federated TRPO'. Appendix A.2 defines the experimental setup by equating a single network seeing stacked observations with FedAvg, and explicitly states 'we are not required to implement the exact flow of FL'. Hence every 'Federated TRPO' performance number is, by the paper's own stipulation, a centralized single-network number. The conclusion 'Federated TRPO performed on par with a tuned federated PPO' therefore reduces to 'our centralized stacked-network TRPO performed on par with our centralized stacked-network PPO'—the very experiments that were run—so the 'federated' label carries no independent empirical content.
full rationale
Strictly, this paper contains no mathematical derivation chain whose outputs are equivalent to its inputs by construction. The 'known optimal policy' is an external benchmark built from the synthetic-data rule, not fitted from the agents' rewards, and the comparisons to tuned PPO, while internally inconsistent with the 'on par' sentence (Tables 1-2 show TRPO eval rewards around -0.47 vs PPO around -0.14), are an accuracy/consistency problem, not a circular one. The one genuinely self-referential step is Appendix A.2, where the paper stipulates that passing stacked observations to one network 'has the same effect as doing a FedAvg'. This is load-bearing for every 'Federated TRPO' statement: the federated label is attached by definition rather than by running a distributed protocol, so results labeled 'federated' are, by the paper's own equivalence, results about a centralized single network. I count this as a minor self-definitional labeling issue, not as a derivation that is forced by construction; the underlying empirical measurements remain self-contained. No load-bearing self-citations or imported uniqueness theorems appear.
Assumptions & free parameters
free parameters (3)
- Reward weights W_C and W_G =
0.4 and 0.6
- Selling-price margin =
40% of the lowest daily electricity price
- PPO hyperparameters =
not specified in the text
assumptions (4)
- ad hoc to paper Centralized stacking of all agents' observations through one shared network has the same effect as FedAvg.
- domain assumption The synthetically generated solar and load profiles guarantee net-zero feasibility, so the known optimal policy has zero cost and zero emissions.
- domain assumption Batteries do not degrade, so the optimal policy is a simple daily charge and discharge cycle.
- domain assumption The reward function with weights W_C and W_G plus the infeasibility penalty fully captures the real objective.
Cite this review
Pith. "Pith review of Generalising Battery Control in Net-Zero Buildings via Personalised Federated RL." pith.science (2026). https://pith.science/paper/DDMGWCZS
@misc{pith2026241220946,
author = {Pith},
title = {Pith review of: Generalising Battery Control in Net-Zero Buildings via Personalised Federated RL},
year = {2026},
howpublished = {\url{https://pith.science/paper/DDMGWCZS}},
note = {Machine review of arXiv:2412.20946}
}
read the original abstract
This work studies the challenge of optimal energy management in building-based microgrids through a collaborative and privacy-preserving framework. We evaluated two common RL algorithms (PPO and TRPO) in different collaborative setups to manage distributed energy resources (DERs) efficiently. Using a customized version of the CityLearn environment and synthetically generated data, we simulate and design net-zero energy scenarios for microgrids composed of multiple buildings. Our approach emphasizes reducing energy costs and carbon emissions while ensuring privacy. Experimental results demonstrate that Federated TRPO is comparable with state-of-the-art federated RL methodologies without hyperparameter tuning. The proposed framework highlights the feasibility of collaborative learning for achieving optimal control policies in energy systems, advancing the goals of sustainable and efficient smart grids. Our code is accessible \href{https://github.com/Optimization-and-Machine-Learning-Lab/energy_fed_trpo.git}{\textit{this repo}}.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[1]
write newline
" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
- [2]
-
[3]
Bekal, G. U., Ghareeb, A., and Pujari, A. Continual reinforcement learning for hvac systems control: Integrating hypernetworks and transfer learning. arXiv preprint, 2025. URL https://arxiv.org/abs/2503.19212
work page Pith review arXiv 2025
-
[4]
Experiment tracking with weights and biases, 2020
Biewald, L. Experiment tracking with weights and biases, 2020. URL https://www.wandb.com/. Software available from wandb.com
2020
-
[5]
Bou, A., Bettini, M., Dittert, S., Kumar, V., Sodhani, S., Yang, X., Fabritiis, G. D., and Moens, V. Torchrl: A data-driven decision-making library for pytorch, 2023. URL https://arxiv.org/abs/2306.00577
arXiv 2023
-
[6]
Model-predictive control and reinforcement learning in multi-energy system case studies
Ceusters, G., Rodr \' guez, R. C., Garc \' a, A. B., Franke, R., Deconinck, G., Helsen, L., Now \' e , A., Messagie, M., and Camargo, L. R. Model-predictive control and reinforcement learning in multi-energy system case studies. CoRR, abs/2104.09785, 2021. URL https://arxiv.org/abs/2104.09785
work page Pith review arXiv 2021
-
[7]
Cuadrado, N. M., Guill \' e n, R. A. G., and Tak \' a c, M. FRESCO: federated reinforcement energy system for cooperative optimization. In Maughan, K., Liu, R., and Burns, T. F. (eds.), The First Tiny Papers Track at ICLR 2023, Tiny Papers @ ICLR 2023, Kigali, Rwanda, May 5, 2023 . OpenReview.net, 2023 a . URL https://openreview.net/pdf?id=75mWq5j4iso
work page 2023
-
[8]
MAHTM: A Multi-Agent Framework for Hierarchical Transactive Microgrids
Cuadrado, N. M., Guti \' e rrez, R. A., Zhu, Y., and Tak \' a c, M. MAHTM: A multi-agent framework for hierarchical transactive microgrids. CoRR, abs/2303.08447, 2023 b . doi:10.48550/arXiv.2303.08447. URL https://doi.org/10.48550/arXiv.2303.08447
work page Pith review arXiv doi:10.48550/arxiv.2303.08447 2023
Show all 39 references
-
[9]
Heterofl: Computation and communication efficient federated learning for heterogeneous clients
Diao, E., Ding, J., and Tarokh, V. Heterofl: Computation and communication efficient federated learning for heterogeneous clients. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021. URL https://...
2021
-
[10]
Potential of artificial intelligence in reducing energy and carbon emissions of commercial buildings at scale
Ding, C., Ke, J., Levine, M., Zhou, N., et al. Potential of artificial intelligence in reducing energy and carbon emissions of commercial buildings at scale. Nature Communications, 15 0 (5916), 2024. doi:10.1038/s41467-024-50088-4
2024 doi
-
[11]
Gernaat, D. E. H. J., de Boer, H. S., Daioglou, V., Yalew, S. G., M \"u ller, C., and van Vuuren, D. P. Climate change impacts on renewable energy supply. Nature Climate Change, 11 0 (2): 0 119--125, Feb 2021. ISSN 1758-6798. doi:10.1038/s41558-020-00949-9. URL https://doi.org...
2021 doi
-
[12]
C., Klein, T., and Nabi, M
Geyer, R. C., Klein, T., and Nabi, M. Differentially private federated learning: A client level perspective. CoRR, abs/1712.07557, 2017. URL http://arxiv.org/abs/1712.07557
2017 arXiv
- [13]
-
[14]
Buildings -- energy system, 2024
International Energy Agency . Buildings -- energy system, 2024. URL https://www.iea.org/energy-system/buildings. Accessed: 7 June 2025
2024
-
[15]
Climate change synthesis report
IPCC. Climate change synthesis report. https://www.ipcc.ch/report/ar6/syr/downloads/report/IPCC_AR6_SYR_SPM.pdf, 2023. (Accessed on 09/13/2024)
2023
-
[16]
B., Avent, B., Bellet, A., and et al
Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., and et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14 0 (1--2): 0 1--210, 2021. doi:10.1561/2200000083
2021 doi
-
[17]
and Langford, J
Kakade, S. and Langford, J. Approximately optimal approximate reinforcement learning. In Proceedings of the Nineteenth International Conference on Machine Learning, pp.\ 267--274, 2002
2002
-
[18]
U., and Jaggi, M
Koloskova, A., Stich, S. U., and Jaggi, M. Decentralized stochastic optimization and gossip algorithms with compressed communication. In Chaudhuri, K. and Salakhutdinov, R. (eds.), Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019,...
2019
-
[19]
Privacy-preserving energy management of a shared energy storage system for smart buildings: A federated deep reinforcement learning approach
Lee, S., Xie, L., and Choi, D.-H. Privacy-preserving energy management of a shared energy storage system for smart buildings: A federated deep reinforcement learning approach. Sensors, 21 0 (14): 0 4898, 2021. doi:10.3390/s21144898
2021 doi
-
[20]
McMahan, B., Moore, E., Ramage, D., Hampson, S., and y Arcas, B. A. Communication-efficient learning of deep networks from decentralized data. In Singh, A. and Zhu, X. J. (eds.), Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTAT...
2017
-
[21]
Review and evaluation of multi-agent control applications for energy management in buildings
Michailidis, P., Michailidis, I., and Kosmatopoulos, E. Review and evaluation of multi-agent control applications for energy management in buildings. Energies (19961073), 17 0 (19), 2024
2024
-
[22]
R., Dey, S., and Henze, G
Nagy, Z., V\' a zquez-Canteli, J. R., Dey, S., and Henze, G. The citylearn challenge 2021. In Proceedings of the 8th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation, BuildSys '21, pp.\ 218–219, New York, NY, USA, 2021. Associa...
2021
-
[23]
Grid integration of zero net energy communities
Narayanamurthy, R., Handa, R., Tumilowicz, N., Herro, C., and Shah, S. Grid integration of zero net energy communities. ACEEE Summer Study Energy Effic. Build, 2016
2016
-
[24]
Real-world challenges for multi-agent reinforcement learning in grid-interactive buildings
Nweye, K., Liu, B., Stone, P., and Nagy, Z. Real-world challenges for multi-agent reinforcement learning in grid-interactive buildings. Energy and AI, 10: 0 100202, 2022 a
2022
-
[25]
The citylearn challenge 2022: Overview, results, and lessons learned
Nweye, K., Nagy, Z., Mohanty, S., Chakraborty, D., Sankaranarayanan, S., Hong, T., Dey, S., Henze, G., Drgona, J., Lin, F., et al. The citylearn challenge 2022: Overview, results, and lessons learned. NeurIPS 2022 Competition Track, pp.\ 85--103, 2022 b
2022
-
[26]
Merlin: Multi-agent offline and transfer learning for occupant-centric energy flexible operation of grid-interactive communities using smart meter data and citylearn
Nweye, K., Sankaranarayanan, S., and Nagy, Z. Merlin: Multi-agent offline and transfer learning for occupant-centric energy flexible operation of grid-interactive communities using smart meter data and citylearn. CoRR, 2023
2023
-
[27]
Citylearn v2: energy-flexible, resilient, occupant-centric, and carbon-aware management of grid-interactive communities
Nweye, K., Kaspar, K., Buscemi, G., Fonseca, T., Pinto, G., Ghose, D., Duddukuru, S., Pratapa, P., Li, H., Mohammadi, J., Lino Ferreira, L., Hong, T., Ouf, M., Capozzoli, A., and Nagy, Z. Citylearn v2: energy-flexible, resilient, occupant-centric, and carbon-aware management o...
2024
-
[28]
Perera, A. T. D., Nik, V. M., Chen, D., Scartezzini, J.-L., and Hong, T. Quantifying the impacts of climate change and extreme climate events on energy systems. Nature Energy, 5 0 (2): 0 150--159, Feb 2020. ISSN 2058-7546. doi:10.1038/s41560-020-0558-0. URL https://doi.org/10....
2020 doi
-
[29]
L., Kaack, L
Rolnick, D., Donti, P. L., Kaack, L. H., Kochanski, K., Lacoste, A., Sankaran, K., Ross, A. S., Milojevic - Dupont, N., Jaques, N., Waldman - Brown, A., Luccioni, A., Maharaj, T., Sherwin, E. D., Mukkavilli, S. K., K \" o rding, K. P., Gomes, C. P., Ng, A. Y., Hassabis, D., Pl...
1906 arXiv
-
[30]
Proximal policy optimization algorithms, 2017
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms, 2017. URL https://arxiv.org/abs/1707.06347
2017 arXiv
-
[32]
Federated reinforcement learning for sustainable and cost-efficient energy management
Sievers, J., Henrich, P., Beichter, M., Mikut, R., Hagenmeyer, V., Blank, T., and Simon, F. Federated reinforcement learning for sustainable and cost-efficient energy management. Energy and AI, 21: 0 100521, 2025 b . doi:10.1016/j.egyai.2025.100521
2025
-
[33]
H., Zhang, N., Li, F., Chen, T., and Cao, H
Su, Z., Wang, Y., Luan, T. H., Zhang, N., Li, F., Chen, T., and Cao, H. Secure and efficient federated learning for smart grid with edge-cloud collaboration. IEEE Trans. Ind. Informatics , 18 0 (2): 0 1333--1344, 2022. doi:10.1109/TII.2021.3095506. URL https://doi.org/10.1109/...
2022
-
[34]
Ton, D. T. and Smith, M. A. The u.s. department of energy's microgrid initiative. The Electricity Journal, 25 0 (8): 0 84--94, 2012. ISSN 1040-6190. doi:https://doi.org/10.1016/j.tej.2012.09.013. URL https://www.sciencedirect.com/science/article/pii/S1040619012002254
2012 doi
-
[35]
K., Balis, J
Towers, M., Kwiatkowski, A., Terry, J. K., Balis, J. U., de Cola, G., Deleu, T., Goulão, M., Kallinteris, A., Krimmel, M., KG, A., Perez-Vicente, R., Pierré, A., Schulhoff, S., Tai, J. J., Tan, H. J. S., and Younis, O. G. Gymnasium: A Standard Interface for Reinforcement Learn...
-
[36]
R., K \" a mpf, J
V \' a zquez - Canteli, J. R., K \" a mpf, J. H., Henze, G., and Nagy, Z. Citylearn v1.0: An openai gym environment for demand response with deep reinforcement learning. In Proceedings of the 6th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, a...
2019
-
[37]
Split learning for health: Distributed deep learning without sharing raw patient data
Vepakomma, P., Gupta, O., Swedish, T., and Raskar, R. Split learning for health: Distributed deep learning without sharing raw patient data. CoRR, abs/1812.00564, 2018. URL http://arxiv.org/abs/1812.00564
2018 arXiv
-
[38]
End-use load profiles for the u.s
Wilson, Eric, Parker, Andrew, Fontanini, Anthony, Present, Elaina, Reyna, Janet, Adhikari, Rajendra, Bianchi, Carlo, CaraDonna, Christopher, Dahlhausen, Matthew, Kim, Janghyun, LeBar, Amy, Liu, Lixi, Praprost, Marlena, White, Philip, Zhang, Liang, DeWitt, Peter, Merket, Noel, ...
2021
-
[39]
R., Ning, L., and Singhal, K
Yuan, H., Morningstar, W. R., Ning, L., and Singhal, K. What do we mean by generalization in federated learning? In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net, 2022. URL https://openreview.net/fo...
2022
-
[40]
J., Qin, Y., et al
Zheng, D., Tong, D., Davis, S. J., Qin, Y., et al. Climate change impacts on the extreme power shortage events of wind–solar supply systems worldwide during 1980–2022. Nature Communications, 15 0 (5225), 2024. doi:10.1038/s41467-024-48966-y
1980 doi
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.