On the Suboptimality of Thompson Sampling in High Dimensions

Authors: Raymond Zhang, Richard Combes

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

Reproducibility Variable Result LLM Response
Research Type Experimental We complement our theoretical results with numerical results and show that in practice TS indeed can perform very poorly in some high dimensional situations.
Researcher Affiliation Academia Raymond Zhang ENS Paris Saclay, Département de Mathématiques, Gif-sur-Yvette, France raymond.zhang@ens-paris-saclay.fr Richard Combes Centrale-Supelec, Laboratoire des Signaux et Systèmes (L2S), Gif-sur-Yvette, France richard.combes@centralesupelec.fr
Pseudocode Yes Figure 1: TS for combinatorial semi-bandits with Bernoulli rewards and uniform prior. Prior knowledge to the learner: combinatorial set X {0, 1}d, function f : X [0, 1]d R For t = 1, ..., T: 1. The learner computes statistics A(t) = Pt 1 s=1 Z(s) x(s) and B(t) = Pt 1 s=1 x(s) Z(s) x(s) 2. The learner draws V (t) with V1(t), ..., Vd(t) independent and Vi(t) Beta(Ai(t) + 1, Bi(t) + 1) 3. The learner chooses decision x(t) arg maxx X {f(x, V (t))} 4. The environment draws Z(t) with Z1(t), ..., Zd(t) independent and Zi(t) Ber(θi) 5. The learner observes Z(t) x(t) and receives the reward f(x(t), Z(t)) Performance metric: expected regret R(T, θ) = T(maxx X {Ef(x, Z(t))}) PT t=1 Ef(x(t), Z(t))
Open Source Code No The paper does not provide any statement or link indicating the availability of open-source code for the described methodology.
Open Datasets No The paper defines the parameters for the problem instances (e.g., 'θi = 1 if 1 i d/2 and θi = 1m otherwise') and generates data based on Bernoulli distributions, but does not refer to or provide access information for a publicly available or open dataset.
Dataset Splits No The paper conducts numerical experiments but does not provide specific details on training, validation, or test dataset splits.
Hardware Specification No The paper does not provide any specific hardware details such as GPU/CPU models, memory, or detailed computer specifications used for running the experiments.
Software Dependencies No The paper does not specify any software names with version numbers used for the implementation or experiments.
Experiment Setup Yes Unless specified otherwise we use 1000 independent sample paths for averaging, and 95% confidence intervals are presented on the plots. Due to limited space, we solely consider the linear case, which is the most often considered in the literature. These experiments are averaged over 40 sample paths due to computational limits. On figures 6a and 6b we present the regret as a function of m for the set of paths X p, m = 0.1, T = 4.104 and the set of matchings X m, m = 0.05, T = 4.104 respectively.