Best of Both Worlds Guarantees for Smoothed Online Quadratic Optimization
Authors: Neelkamal Bhuyan, Debankur Mukherjee, Adam Wierman
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To further explore SOQO in stochastic and adversarial settings, we conduct empirical experiments to evaluate the performance of our algorithms in a range of environments. Our experiments fall into two main categories: first, purely stochastic experiments where we provide empirical evidence for two primary claims the inferior performance of ROBD in comparison to LAI(γ) and LAI, and the robustness of our algorithms to distribution shifts and variations in tail behavior (light or heavy). Second, we delve into mixed adversarial-stochastic experiments that combine elements of both adversarial and stochastic scenarios, examining how our algorithms perform in this hybrid setting. |
| Researcher Affiliation | Academia | 1H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA, USA 2Department of Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA, USA. Correspondence to: Neelkamal Bhuyan <nbhuyan3@gatech.edu>. |
| Pseudocode | Yes | Algorithm 1 Lazy Adaptive Interpolation Algorithm 2 Lazy Adaptive Interpolation (γ) |
| Open Source Code | No | The paper does not provide any explicit statement or link indicating that its source code for the described methodology is publicly available. |
| Open Datasets | No | The paper describes how it generates synthetic data for experiments, specifying parameters and distributions (e.g., 'action space Rd', 'eigenvalues selected from one of three sequences', 'sequence of minimizers {vt}t', 'light tail distributions', 'heavy-tail distributions', 'IID sequence of increments'), but it does not use or provide access information for a pre-existing public dataset. |
| Dataset Splits | No | The paper describes methods for generating synthetic data and setting up experimental scenarios (e.g., 'split the horizon into five subsets', 'adversarial percentage (p)'), but it does not refer to traditional training/validation/test dataset splits for a fixed dataset, as its data is generated on the fly for simulations. |
| Hardware Specification | No | The paper does not provide any specific details regarding the hardware (e.g., GPU/CPU models, memory) used to run the experiments. |
| Software Dependencies | No | The paper does not mention any specific software or library dependencies with version numbers used for the experiments. |
| Experiment Setup | Yes | In all of our experiments, we maintain a consistent action space R10 and employ the matrix A with eigenvalues selected from one of three sequences: 0.3i 9 i=0, 0.45i 9 i=0, or 0.5i 9 i=0. The sequence of minimizers, denoted {vt}t, is adjusted according to the specific type of experiment under consideration. Our analysis involves plotting the expected total regret (relative to LAI) for both 0.3i 9 i=0 and 0.5i 9 i=0 over various horizon values T {1, 2, . . . , 100}. We present the results as the sample mean derived from N = 1000 runs, accompanied by the 95th percentile for added clarity. In this series of experiments, we introduce adversarial minimizers into a martingale minimizer sequence, with the extent of adversarial influence determined by the parameter known as the adversarial percentage denoted (p). This parameter spans from 0 (indicating a fully stochastic scenario) to 100 (representing a fully adversarial scenario). |