Additive Causal Bandits with Unknown Graph

Authors: Alan Malek, Virginia Aglietti, Silvia Chiappa

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This section presents an empirical evaluation of the MODL algorithm on a collection of randomly generated causal additive models.1 Additional experiments studying the effect of graph structure and the sensitivity to the additive outcome assumption s violation can be found in Appendix B.
Researcher Affiliation Industry Alan Malek 1 Virginia Aglietti 1 Silvia Chiappa 1 1DeepMind, London, UK.
Pseudocode Yes Pseudocode is provided in Algorithm 1. Pseudocode and a complexity bound are provided in Appendix A.
Open Source Code Yes 1Code has been released at https://github.com/ deepmind/additive_cbug.
Open Datasets No We performed the evaluation on randomly sampled structural causal models (SCMs) generated as followed. The causal graph, excluding Y , was a sampled directed acyclic graph from the Erd os-R enyi model with the degree 3 and K 1 variables. We randomly choose set of variables of size PY as the parents of Y , then each variable topologically greater than Y is independently set as a child to Y with probability .5.
Dataset Splits No The paper describes generating its own data and running experiments on it but does not specify any training/validation/test dataset splits, percentages, or references to predefined splits.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., library names with specific versions like 'PyTorch 1.9' or 'Python 3.8').
Experiment Setup Yes Experimental Set-up. We performed the evaluation on randomly sampled structural causal models (SCMs) generated as followed. The causal graph, excluding Y , was a sampled directed acyclic graph from the Erd os-R enyi model with the degree 3 and K 1 variables. We randomly choose set of variables of size PY as the parents of Y , then each variable topologically greater than Y is independently set as a child to Y with probability .5. To sample the conditional probability distributions, we chose Mk uniformly between specified upper and lower bounds and generated the conditional probability distribu- tion for each Xk by sampling p(Xk = j | pa(Xk) = x) Beta(2, 5) independently for all j and x. Finally, we generated fk(j) = BW j k, with B = 5 and W j k sampled i.i.d. from Beta(2, 5) and set η to a standard normal variable. If Xj had Y as a parent, we used the same construction but with Y rounded to an integer. Results. Using ϵ = 1/2 and δ = .1 for our (ϵ, δ)-PAC criterion, we considered a variety of settings of PY , K, and upper and lower bounds for Mk. Each point all graphs corresponds to the average over 20 different SCMs sampled using the process described above and 50 independent runs of the methods on independently generated data.