Pure Exploration and Regret Minimization in Matching Bandits

Authors: Flore Sentenac, Jialin Yi, Clement Calauzenes, Vianney Perchet, Milan Vojnovic

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we present numerical results in Section 5. They validate the tightness of our results and demonstrate competitiveness and performance gains obtained by our proposed algorithms over some state-of-the-art baseline algorithms.
Researcher Affiliation Collaboration 1CREST, ENSAE Paris, Palaiseau, France 2London School of Economics, London, UK 3Criteo AI Lab, Paris, France 4CREST, ENSAE Paris, Palaiseau, France & Criteo AI Lab, Paris, France.
Pseudocode Yes Algorithm Sketch 1 PAIR-ELIM, Algorithm Sketch 2 PAIR-ELIM-MONO, Algorithm Sketch 3 MATCHING-ID, and Algorithm 4 SIMPLE-ADAPTIVE-MATCHING are all present in the paper.
Open Source Code Yes All the code used for obtaining the results in this section is available from this public Gitlab repository: [anonymized]
Open Datasets No The experiments are conducted on 'problem instances' generated based on specified parameters (N, u1, etc.) and Bernoulli variables, rather than a named, publicly accessible dataset. There is no link, DOI, or formal citation provided for a public dataset.
Dataset Splits No The paper defines 'problem instances' for its simulations but does not describe specific train, validation, or test dataset splits (e.g., percentages or sample counts) for reproducibility of data partitioning.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper does not specify any software dependencies or libraries with version numbers (e.g., 'Python 3.8, PyTorch 1.9').
Experiment Setup Yes We consider the bipartite case with N = M. Each problem instance is defined by a tuple (N, u1, ), where 0 u1 1, and assuming that u1 = v1. and The results are shown in Figure 1. We have further evaluated the effects of varying u1 which is discussed in Appendix A.1. and We consider a set of problem instances, each is defined by a tuple (N, ), where 0 < (N 1) 1 and is the gap between parameter values of adjacent matched pairs in the optimum matching, so that = min. The parameter values are defined by u2i 1 = u2i = (N i) for i [N]. and The results shown in Figure 2 suggest that the cumulative regret of SIMPLE-ADAPTIVE-MATCHING scales as (1/ min)N log(N) as established in Theorem 4.3. We next compare the performance of our algorithm and ESCB. We consider problem instances defined by a tuple (N, µ, ) where µ is the mean of parameter values and is the gap between parameter values. The values of item parameters are set as u2i 1 = u2i = µ+(N+1 2i) /2 [0, 1] for i [N]. Such problem instances allow us to vary µ while keeping other parameters fixed; we fix N = 4 and = 0.1.