Model Selection in Contextual Stochastic Bandit Problems

Authors: Aldo Pacchiano, My Phan, Yasin Abbasi Yadkori, Anup Rao, Julian Zimmert, Tor Lattimore, Csaba Szepesvari

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiment (Figure 1). Let d = 2. Consider a contextual bandit problem with k = 50 arms, where each arm j has an associated vector aj 2 Rd sampled uniformly at random from [0, 1]d. We consider two cases: (1) For a 2 Rd sampled uniformly at random from [0, 1]d, reward of arm j at time t is a> j + t, where t N(0, 1), and (2) There are k parameters µj for j 2 [k] all sampled uniformly at random from [0, 10], so that the reward of arm j at time t is sampled from N(µj, 1). We use CORRAL with learning rate = 2 p T d and UCB and Lin UCB as base algorithm. In case (1) Lin UCB performs better while in case (2) UCB performs better. Each experiment is repeated 500 times.
Researcher Affiliation Collaboration Aldo Pacchiano UC Berkeley My Phan University of Massachusetts Yasin Abbasi-Yadkori Julian Zimmert Google Research Tor Lattimore Csaba Szepesvári DeepMind and University of Alberta
Pseudocode Yes Algorithm 1 Master Algorithm Input: Base Algorithms {Bj}M j=1 for t = 1, , T do Play base jt. Receive feedback rt = rt,jt from Bjt Update itself using rt end for
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 describes synthetic data generation setups for its experiments, such as actions sampled uniformly at random from a distribution or Bernoulli arms, but it does not provide concrete access information for any publicly available or open dataset.
Dataset Splits No The paper describes experimental setups with synthetic data but does not specify any train/validation/test dataset splits, as it does not rely on pre-existing benchmark datasets with defined splits.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments.
Software Dependencies No The paper discusses algorithms and mathematical formulations but does not mention specific software dependencies or their version numbers required for replication.
Experiment Setup Yes Experiment (Figure 1). Let d = 2. Consider a contextual bandit problem with k = 50 arms... We use CORRAL with learning rate = 2 p T d and UCB and Lin UCB as base algorithm... Experiment (Figure 2)... We take T = 50, 000, = 20/T and s to lie on a geometric grid in [1, 2T].