Piecewise-Stationary Bandits with Knapsacks

Authors: Xilin Zhang, Wang Chi Cheung

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

Reproducibility Variable Result LLM Response
Research Type Experimental We run numerical experiments on a single-resource problem where L 2, T 20000 (each stationary piece has 10000 rounds), K t1, 2u, B 9360 and we set α e for our algorithms. The rewards and resource consumption in all rounds are uniformly distributed within a r 0.2, 0.2s range from their mean values. We compare the performance of IRES-CM with Immorlica et al. (2019) s algorithm and Zhou et al. (2008) s algorithm. Recall that Immorlica et al. (2019) focus on an adversarial Bwk problem and achieves a CR w.r.t. a static benchmark. Zhou et al. (2008) study a full-feedback adversarial setting and achieves a CR w.r.t. a single best arm benchmark. In Figure 1, each curve represents the average cumulative reward over 10 simulations, and the shaded area around each curve marks the variance over the simulations.
Researcher Affiliation Academia Xilin Zhang Department of ISEM National University of Singapore Singapore, 117578 zhangxilin@u.nus.edu Cheung Wang Chi Department of ISEM National University of Singapore Singapore, 117578 isecwc@nus.edu.sg
Pseudocode Yes Algorithm 1 Inventory REServing with deterministic input (IRES) ... Algorithm 2 Inventory REServing with Change Monitoring (IRES-CM)
Open Source Code No Our data are numerically generated and codes can be provided.
Open Datasets No The paper generates its own data for numerical experiments: 'The rewards and resource consumption in all rounds are uniformly distributed within a r 0.2, 0.2s range from their mean values.'
Dataset Splits No The paper describes simulation parameters and runs, but does not specify a train/validation/test dataset split as it uses generated data in an online learning setting.
Hardware Specification Yes Finally, our experiments are run on a Surface Pro 7 with an i5-1035G4 processor.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes We run numerical experiments on a single-resource problem where L 2, T 20000 (each stationary piece has 10000 rounds), K t1, 2u, B 9360 and we set α e for our algorithms.