Linear Causal Bandits: Unknown Graph and Soft Interventions

Authors: Zirui Yan, Ali Tajer

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we assess the regret performance of GA-LCB. As the most relevant existing approaches, we compare the regret of our algorithm to those of Lin SEM-UCB [15] and GCB-UCB [18], which are designed for causal bandits with soft interventions. ... Figure 2: Cumulative regret with L = 2.
Researcher Affiliation Academia Zirui Yan Rensselaer Polytechnic Institute yanz11@rpi.edu Ali Tajer Rensselaer Polytechnic Institute tajer@ecse.rpi.edu
Pseudocode Yes Algorithm 1 Graph-Agnostic Linear Causal Bandit: Structure Learning (GA-LCB-SL) ... Algorithm 2 Graph-Agnostic Linear Causal Bandit: Intervention Design (GA-LCB-ID)
Open Source Code Yes See the details in Appendix A and the codes in https://github.com/ Zirui Yan/Linear-Causal-Bandit-Unknown-Graph.
Open Datasets No Causal graph. We consider the hierarchical graph illustrated in Figure 1. This graph consists of (L + 1) layers, with the first L layers having d nodes. ... The noise terms {ϵi : i [N]} are set to be drawn from the uniform distribution Unif(0, 1).
Dataset Splits No The paper describes generating synthetic data based on a hierarchical graph and setting parameters (e.g., noise terms, weight matrices, exploration times T1 and T2), but does not define traditional train/validation/test dataset splits.
Hardware Specification Yes The experiment was conducted using 2 CPUs from Mac Mini 2023.
Software Dependencies No The paper does not list specific versions for any software components, libraries, or programming languages (e.g., Python, PyTorch) that would be necessary for reproducible replication.
Experiment Setup Yes We set the non-zero elements in the observational and interventional weights matrix to 1 and 0.5, respectively. ... We set T1 = T2 = 500 for experiments with L = 2, T1 = T2 = 1000 for experiments with L = 4 and L = 6. ... We observe that setting λ = α = 0.1 yields reasonable performance.