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..
Beyond $\tilde{O}(\sqrt{T})$ Constraint Violation for Online Convex Optimization with Adversarial Constraints
Authors: Abhishek Sinha, Rahul Vaze
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results, presented in Section 8.7 of the Appendix, qualitatively support the expected variations of Regret and CCV as a function of the parameter β. 8.7 Experiments Problem instance: We consider the CONSTRAINED EXPERT problem on a synthetic dataset with N = 20 experts over a horizon of length T = 5000. Two experts are designated as special: the best feasible expert, denoted by E , and the best unconstrained expert, denoted by UE . The expert E is feasible, with i.i.d. costs drawn from a distribution with mean f E = 0.21, and zero constraint violation on all rounds, i.e., g E = 0.0. In contrast, UE is infeasible, with i.i.d. costs and constraint violations having a smaller mean f UE = 0.11 and higher average constraint violation g UE = 0.91, respectively. The remaining experts incur i.i.d. random costs with mean f = 0.41 plus a zero-mean periodic component over time, and their constraint violations are i.i.d. with mean g = 0.6. Additionally, two more experts, DE1 and DE2, distinct from both E and UE , are made feasible by setting their constraint violations to zero on all rounds. Table 2 summarizes the parameters used in the experiments. The experiments have been run on a quad-core CPU with 8 GB RAM. The source code has been made publicly available [Sinha, 2025]. |
| Researcher Affiliation | Academia | Abhishek Sinha Rahul Vaze School of Technology and Computer Science Tata Institute of Fundamental Research Mumbai 400005, India EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 presents our proposed online policy for the CONSTRAINED EXPERT problem. Algorithm 2 COCO Algorithm for Convex Cost and Constraints Algorithm 3 COCO Algorithm for Smooth and Convex Cost and Constraints Algorithm 4 Adaptive Hedge Algorithm for the EXPERT problem with Unbounded Losses |
| Open Source Code | Yes | The code has been made publicly available [Sinha, 2025]. 8.7 Experiments ...The source code has been made publicly available [Sinha, 2025]. |
| Open Datasets | No | We consider the CONSTRAINED EXPERT problem on a synthetic dataset with N = 20 experts over a horizon of length T = 5000. Two experts are designated as special: the best feasible expert, denoted by E , and the best unconstrained expert, denoted by UE . The expert E is feasible, with i.i.d. costs drawn from a distribution with mean f E = 0.21, and zero constraint violation on all rounds, i.e., g E = 0.0. In contrast, UE is infeasible, with i.i.d. costs and constraint violations having a smaller mean f UE = 0.11 and higher average constraint violation g UE = 0.91, respectively. The remaining experts incur i.i.d. random costs with mean f = 0.41 plus a zero-mean periodic component over time, and their constraint violations are i.i.d. with mean g = 0.6. Additionally, two more experts, DE1 and DE2, distinct from both E and UE , are made feasible by setting their constraint violations to zero on all rounds. Table 2 summarizes the parameters used in the experiments. |
| Dataset Splits | No | The paper uses a synthetic dataset described in Section 8.7. It details the generation of expert costs and constraint violations but does not specify any training, test, or validation splits for this dataset. |
| Hardware Specification | Yes | The experiments have been run on a quad-core CPU with 8 GB RAM. |
| Software Dependencies | No | The paper states that the source code has been made publicly available [Sinha, 2025], but it does not specify any particular software dependencies with version numbers (e.g., Python version, library names, and their versions) used in the implementation. |
| Experiment Setup | Yes | Problem instance: We consider the CONSTRAINED EXPERT problem on a synthetic dataset with N = 20 experts over a horizon of length T = 5000. ... Table 2 summarizes the parameters used in the experiments. Parameter settings for generating cost and constraint vectors for N = 20 experts. Summary of the results. Parts (A) and (B) of Figure 2 compare the performance of our proposed adaptive Hedge-based policy (Algorithm 1) with β = 0.75 with that of the Online Gradient Descent (OGD)-based policy proposed by Sinha and Vaze [2024, Algorithm 1]. |