acceptodds
Withdrawn from ICLR 2027

FULCRUM: A Preference-Conditioned Hierarchical Graph Reinforcement Learning for Sustainable Multi-Data-Center Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Geo-distributed data-center fleets must trade off electricity costs, carbon emissions, water use, and quality of service (QoS) across sites with varying prices, grids, and weather conditions. Learned controllers for this problem are trained for a single objective weighting and a single fleet size, so any change in operator priorities or footprint requires retraining. We present , a single preference-conditioned graph policy: a Graph Pattern Machine whose objective-aligned attention heads are feature-wise linear modulation (FiLM) conditioned on the operator preference \pref, followed by cross-attention from data-center nodes to preference-scaled objective tokens. All parameters are shared across nodes and edges, so the same weights act on any fleet. Under a budget-matched protocol with paired tests over 50 evaluation episodes, repeated across three training seeds, attains the best mean rank across five operator preferences among ten controllers (), matches or exceeds every preference-conditioned baseline across all seven held-out preferences never used in training ( practical wins, no losses against preference-conditioned baselines), and at one-hot preferences reaches the lowest cost, carbon, water, and QoS penalty of any baseline. It steers the fleet at inference time from water to cost relative to its balanced operating point and, trained once at  data centers, outperforms a Green-DCC controller retrained from scratch at each of , , , and data centers by – reward, where preference-conditioned MLP and GNN baselines cannot be loaded at all. One training run replaces five per-preference specialists at lower compute for a reward cost. We also report where does not lead: at intermediate preference weights, conditioned baselines use – less water, one of three seeds emits more carbon at the carbon corner, and the cross-attention and FiLM components are not individually significant at this budget.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Sustainability, Optimization, Graph Pattern Machines

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_YzZcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes FULCRUM, a data center scheduling method based on preference-conditional graph neural networks, which can dynamically balance electricity costs, carbon emissions, water consumption, and service quality through a single model.

Critical strengths

None

Critical weaknesses

The paper is poorly organized. Issues include, but are not limited to, the following: 1) The introduction lists three limitations, yet it remains unclear how the proposed method specifically addresses them; 2) The calculation methods for the various metrics of "Green-DCC" in Table 1 are not explained; 3) The comparative methods are outdated, lacking the latest baselines; 4) The proposed method lacks novelty, amounting to little more than the application of an attention mechanism.

Questions for authors

See Weaknesses

1 comment

The authors have not responded. Therefore, I will maintain my original score.

Reviewer_8b8BOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the control of geo-distributed data centers and proposes FULCRUM, a preference-conditioned hierarchical graph reinforcement learning framework that enables zero-shot generalization.

Critical strengths
  1. The paper studies a practically relevant problem of controlling geo-distributed data center fleets.
  2. The framework combines preference adaptation with transfer across fleet sizes.
Critical weaknesses
  1. The broader methodological contribution beyond the data center application is not clear. The main capabilities of FULCRUM mainly follow from existing preference-conditioned RL and graph-based policies. It is not clear what special aspects of the studied problem can lead to new methodological insights that can be applied to a set of previously underexplored problems.

  2. It seems from Table 5 that removing either FiLM or preference cross-attention does not have a significant impact on the algorithm's performance. This makes the readers wonder about the values of these components. Are they really necessary in the proposed algorithm? This seems to undermine the contribution.

  3. Eq. (3) requires clarification. First, it seems that βk(ω)\beta_k(\omega) does not depend on neighbor jj and hence does not impact the softmax outcomes. Then what is the purpose of having this term? Second. It seems that the heads are objective-aligned, but it is not clear how the heads learn the patterns associated with the corresponding objectives.

Questions for authors
  1. What is the value of the components in this method (e.g., FiLM and preference cross-attention), especially given the ablation studies?
  2. Table 8 shows that the trade-off between different objective functions can be rather extreme. How can the proposed framework ensure a reasonable trade-off in practice, such that the resulting control decisions remain operationally acceptable?
Reviewer_qcatOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a preference-conditioned graph reinforcement learning framework for sustainable control of geographically distributed data centers, considering four objectives: electricity cost, carbon emissions, water consumption, and quality of service. To handle varying numbers of data centers, the method represents the system as a directed graph and employs a graph neural network with shared node and edge level operations. To adapt the policy to different objective preferences, the authors introduce feature-wise linear modulation (FiLM) and preference cross-attention into the network.

Critical strengths
  1. The paper addresses a practical multi-objective control problem in geographically distributed data centers. The proposed graph-based policy can handle different numbers of data centers without retraining, while allowing operators to adjust the tradeoffs among objectives.

  2. The experiments include evaluations on both synthetic and real-world workload traces and tests of generalization across different fleet sizes.

Critical weaknesses
  1. The proposed method is somewhat incremental, since it only combines a graph neural network and the FiLM technique as the policy in multi-objective reinforcement learning to solve the multi-data-center problem. The authors should at least show that the proposed method can be extended to other relevant problems, or that this problem has strong connections to other important problems, such that the proposed method has the potential to be extended to other domains.

  2. Multi-objective reinforcement learning has advanced in recent years; thus, it would be valuable to apply at least two existing popular MORL methods to this problem.

  3. The FiLM technique is not the only way to incorporate the preference vector into a neural network. Other classical methods for integrating preference vectors in Pareto front learning should also be considered.

Questions for authors

Please see the Weaknesses.

Reviewer_7zNrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FULCRUM, a preference-conditioned graph reinforcement learning framework for geo distributed datacenter scheduling considering four objectives: electricity costs, carbon emissions, water consumption, and service quality using one model without requiring retraining for different preferences.

Critical strengths

1, The paper proposes a single policy that handles different operator preferences without retraining. 2 , Designed for geo-distributed fleets.

Critical weaknesses

1, The energy savings from avoiding retraining seem small compared with the performance loss. 2 , the use of older cpu workload traces may limit how the results apply to modern AI data centers. 3, figure 1 does not fully match the architecture described in the methodology which needs clarification.

Questions for authors

see Weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.