Sparsity-Agnostic Linear Bandits with Adaptive Adversaries

Authors: Tianyuan Jin, Kyoungseok Jang, Nicolò Cesa-Bianchi

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments on synthetic data, Ada Lin UCB performs better than OFUL. In this section we describe some experiments we performed on synthetic data to verify whether Ada Lin UCB could outperform OFUL in a model selection task.
Researcher Affiliation Academia Tianyuan Jin Department of Electrical and Computer Engineering National University of Singapore Singapore tianyuan@nus.edu.sg Kyoungseok Jang Dipartimento di Informatica Università degli Studi di Milano Milano, Italy ksajks@gmail.com Nicolò Cesa-Bianchi Università degli Studi di Milano Politecnico di Milano Milano, Italy nicolo.cesa-bianchi@unimi.it
Pseudocode Yes Algorithm 1 Sparse Lin UCB
Open Source Code Yes The code used in the experiments can be found in the following repository: https://github.com/ jajajang/sparsity_agnostic_model_selection.
Open Datasets No The data for our model selection experiments are generated using targets θ with different sparsity levels... We run our experiments with a fixed set of random actions, At = A for all t [T], where |A| = 30 and A is a set of vectors drawn i.i.d. from the unit sphere in R16. The target vector θ is a S-sparse (S = 1, 2, 4, 8, 16) vector whose non-zero coordinates are drawn from the unit sphere in RS. The noise εt is drawn i.i.d. from the uniform distribution over [ 1, 1].
Dataset Splits No The data for our model selection experiments are generated using targets θ with different sparsity levels... Each curve is an average over 20 repetitions with T = 104 where, in each repetition, we draw fresh instances of A and θ .
Hardware Specification Yes Hardware: Lenovo Thinkpad P16s Gen 2 Laptop Type 21HL CPU: 13th Gen Intel(R) Core(TM) i7-1360P 2.20 GHz RAM: 32GB Computation time: total 1338.38 seconds.
Software Dependencies No No specific software dependencies with version numbers are listed in the provided text.
Experiment Setup Yes Number of iterations: T = 104. Number of models: n = 6. Radius of confidence sets: α0 = 0, and αi = 2i log t for i = 1, , 5. Prior distribution {qs}s [6] For _Unif,{qs}s [6] = 1/6. For _Theory, {qs}s [6] = C/64 where C = 63/64 is a normalizing constant. The learning rate of Exp3 was set to ηt = 2 q nt , see [5].