Cooperative Multi-Agent Bandits with Heavy Tails

Authors: Abhimanyu Dubey, Alex ‘Sandy’ Pentland

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments using α-stable densities (L evy, 1925), that admit finite moments only of order < α 2, and we consider α-stable densities where α 1. The α-stable family includes several widely used distributions, such as Gaussian (α = 2, only lighttailed density), L evy (α = 0.5) and Cauchy (α=1).
Researcher Affiliation Academia 1Media Lab and Institute for Data, Systems and Society, Massachusetts Institute of Technology. Correspondence to: Abhimanyu Dubey <dubeya@mit.edu>.
Pseudocode Yes Algorithm 1 DECENTRALIZED MP-UCB; Algorithm 2 CENTRALIZED MP-UCB; Algorithm 3 ONLINE TRIMMED MEAN ESTIMATOR
Open Source Code No The paper does not contain any explicit statements about releasing source code or providing links to code repositories for the described methodology.
Open Datasets Yes We select the p2p-Gnutella04 (Figure 1C) and ego-Facebook (Figure 1D) network structures from the SNAP repository (Leskovec & Sosiˇc, 2016) to experiment with in the real-world setting.
Dataset Splits No The paper describes generating random graphs ('Erdos-Renyi (ER) (p = 0.7) and Barabasi-Albert (BA) (m = 5) random graph families') and sampling subgraphs from real-world networks ('subgraphs of 500 nodes'). However, it does not specify any explicit train/validation/test dataset splits or cross-validation setups.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory, or cloud instance types) used for running the experiments.
Software Dependencies No The paper mentions using 'the approximate algorithm presented in (Lucas, 2014) that uses the QUBO (Glover & Kochenberger, 2018) solver' but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We set K = 5, α = 1.9 for the standard α-stable density, and sample arm means randomly from the interval [0, 1] for each arm every experiment. We then construct random graphs on 200 agents from the Erdos-Renyi (ER) (p = 0.7) and Barabasi-Albert (BA) (m = 5) random graph families, and compare all three of our algorithms (using the trimmed mean estimator, with γ = diam(G)/2) with the CONSENSUS-UCB and single-agent ROBUST-UCB(Bubeck et al., 2013) algorithms.