Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Fairness-Regularized Online Optimization with Switching Costs
Authors: Pengfei Li, Yuelin Han, Adam Wierman, Shaolei Ren
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 Empirical Results Our empirical study investigates the problem of fair resource provisioning for AI inference services in geographically separated data centers... Table 2: The average costs of different algorithms in the default setting (i.e. u1 = 10, u2 = 1000 and u3 = 3.5). |
| Researcher Affiliation | Academia | Pengfei Li School of Information Rochester Institute of Technology EMAIL Yuelin Han Electrical and Computer Engineering University of California, Riverside EMAIL Adam Wierman Computing & Mathematical Sciences California Institute of Technology EMAIL Shaolei Ren Electrical and Computer Engineering University of California, Riverside EMAIL |
| Pseudocode | Yes | Algorithm 1 Fair OBD: Fairness-regularized Online Balanced Descent 1: Input: Initial κ1, reference function h( ), and learning rate η 2: for t = 1 to T do 3: Receive ft( ) and At 4: Obtain the action xt and auxiliary variable zt by solving the following: min xt Xt,zt Z ft(xt) + λ1d(xt, xt 1) + λ2 2 xt vt 2 + κt Atxt + g(zt) κtzt , (5) where vt = arg minx X ft(x). 5: Obtain a stochastic subgradient of κt: dt = zt At xt 6: Update the dual variable by mirror descent: κt+1 = arg minκ RM dt, κ + 1 ηVh(κ, κt), where Vh(x, y) = h(x) h(y) h(y) (x y) is the Bregman divergence. 7: end for |
| Open Source Code | No | In our empirical study, all information is based on publicly available data. We will make the data and code public upon acceptance of the paper. |
| Open Datasets | Yes | We employ a publicly available inference trace dataset for LLM services on Azure [37]. [37] Microsoft Azure. Azure Public Dataset. https://github.com/Azure/ Azure Public Dataset, 2024. |
| Dataset Splits | Yes | This testing dataset consists of 97 three-day (72-hour) context sequences, created by applying a sliding window across a one-week (168-hour) context sequence. |
| Hardware Specification | Yes | Our experiments are conducted on a Mac Book Air with an M3 chip and 16 GB of memory. |
| Software Dependencies | No | The paper does not explicitly state specific versions of software dependencies (e.g., Python, libraries, frameworks) used for implementing the algorithms. |
| Experiment Setup | Yes | In the default setting, we set q = 1 to map the provisioned computing resource to energy consumption. The weights are u1 = 10 for the regularizer in hitting cost, u2 = 1000 for switching cost and u3 = 3.5 for the long-term fairness cost. We choose the identical maximum capacity for each data center with Mi = 1 and normalize maximum workload traces according to the maximum capacity. We use p = for the lp norm. These values are chosen to ensure that the hitting cost, switching cost, and fairness cost have comparable magnitudes. We set the initial dual variable κ1 = [3]N i=1, λ1 = 1, λ2 = 30 for DMD and Fair OBD. The hyperparameters λ1 and λ2 for ROBD are optimally selected based on Theorem 4 in [18]. For DMD, we use the same set of hyperparameter as Fair OBD, such as the default learning rate η = 10 3, except that DMD ignores the switching cost. |