Smoothed Online Combinatorial Optimization Using Imperfect Predictions

Authors: Kai Wang, Zhao Song, Georgios Theocharous, Sridhar Mahadevan

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, our algorithm shows a significant improvement in cumulative regret compared to other baselines in synthetic online distributed streaming problems.Empirically, we evaluate our algorithm on the online distributed streaming problem motivated from Apache Kafka with synthetic traffic. We compare our algorithm using predictions and dynamic planning windows with various baselines.
Researcher Affiliation Collaboration Kai Wang1*, Zhao Song2, Georgios Theocharous2, Sridhar Mahadevan2 1Harvard University, Cambridge, MA 2Adobe Research, San Jose, CA kaiwang@g.harvard.edu, {zsong, theochar, smahadev}@adobe.com
Pseudocode Yes Algorithm 1: Dynamic Future Planning
Open Source Code No The paper does not provide any explicit statements or links indicating that the source code for the methodology is openly available.
Open Datasets No In our experiment, we use the distributed streaming system problems with synthetic data to compare our algorithm with other baselines.We generate k time series, where each represents the trend of incoming traffic {θt,i}t [T ] of topic i [k] as the cost function parameter. Each time series is generated by a composition of sine waves, an autoregressive process, and a Gaussian process to model the seasonality, trend, and the random process.
Dataset Splits No The paper mentions using “50 historical data” for Gaussian process regression stabilization and “another 100 time steps with hidden incoming traffic to measure the performance,” but it does not specify explicit train/validation/test dataset splits with percentages, absolute counts, or references to predefined splits.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory used for running its experiments.
Software Dependencies No The paper mentions software components like “Apache Kafka”, “Gaussian process regression”, and “mixed integer program”, but it does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes For each instance of the load balancing problem, we assume 50 historical data have been collected a priori to stabilize Gaussian process regression. We run different online algorithms for another 100 time steps with hidden incoming traffic to measure the performance of online algorithms. For each setup, we run 10 independent trials with different random seeds to estimate the average performance.