Dual Mirror Descent for Online Allocation Problems

Authors: Santiago Balseiro, Haihao Lu, Vahab Mirrokni

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 6. Numerical Experiment Here, we present a numerical experiment on proportional matching with high entropy (Section 5.2) to verify our results. Figure 1 plots the regret versus horizon T, from which we can clearly see that the regret grows at the rate of T, which verifies the results in Theorem 1. Figure 2 plots the relative reward (ratio between the reward collected by the online algorithm and the offline optimal) versus horizon T.
Researcher Affiliation Collaboration Santiago Balseiro 1 2 Haihao Lu 2 Vahab Mirrokni 2 ... 1Columbia University, New York, USA 2Google Research, New York, USA.
Pseudocode Yes Algorithm 1 Dual Mirror Descent Algorithm for (1) ... Algorithm 2 Online Dual Mirror Descent Algorithm for Bidding in Repeated Auctions ... Algorithm 3 Online Dual Mirror Descent Algorithm for Proportional Matching Problems with High Entropy
Open Source Code Yes The details of the numerical experiment and data generation are presented in Appendix H, and the code to reproduce the results is in supplementary materials.
Open Datasets Yes The dataset is generated following the procedures stated in Balseiro et al. (2014).
Dataset Splits No The paper describes generating '100,000 samples' and using '20 random datasets with size T uniformly randomly chosen from 100,000 samples' for experiments, but it does not specify explicit train/validation/test splits for these datasets.
Hardware Specification No The paper does not provide any specific hardware details used for running experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers.
Experiment Setup Yes For all experiments, we start from µ0 = 0, utilize h(x) = 1/2 x 2 2 as our reference function (thus the algorithm is dual sub-gradient descent), and choose η = 1/T as the step-size. ... We incorporate the entropy regularizer H(x) to the objective with parameter λ = 0.0002