The glip package [1] provides global optimization-based algorithms for learning (causal) graphical models from data, including Directed Acyclic Graphs (DAGs), Acyclic Directed Mixed Graphs (ADMGs), and chain graphs (CGs). The package relies on mixed-integer programming via GUROBI [2] for accurate, constraint-based graph learning, with advanced support for latent variables and complex causal graphs.

• Learn DAGs, ADMGs, DMGs, CGs directly from observational or interventional data
• Global graph optimization powered by integer programming (requires GUROBI)
• Supports various conditional independence tests (including tests from comets [3], pcalg, and bnlearn)
• Parallel execution for heavy workloads

CRAN release: (Not yet available)

Development version via GitHub (requires R >= 4.1):

# Install remotes if not present
# install.packages("remotes")
remotes::install_github("LucasKook/glip")

The RBGL package (a pcalg dependency) is available through Bioconductor:

if (!require("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install("RBGL")

This package uses GUROBI Optimizer (commercial, free for academic use) for solving mixed-integer problems crucial for global graph learning. You must install , obtain a valid license, and install the gurobi R package by following the instructions in [2].

Learn a causal graph from toy continuous data:

library("glip")

### Set seed for reproducibility
set.seed(1)

### Generate some toy data from: Y <- X -> Z
n <- 500
X <- rnorm(n)
Y <- X + rnorm(n)
Z <- X + rnorm(n)
data <- data.frame(X = X, Y = Y, Z = Z)

# Learn a DAG (using Gaussian CI test from `pcalg`)
result <- learn_graph(data, max_size = 1, mode = "dag", test = "gaussCItest", comets = FALSE)

# Inspect the learned DAG
result$graph

# Inspect the essential graph
result$computed

• "dag": Directed Acyclic Graph (standard causal Bayesian Network)
• "dg": Directed Graph (allows for cycles, used d-separation and hence assumes normal distribution)
• "admg": Acyclic Directed Mixed Graph (allows for hidden confounding)
• "dmg": Directed Mixed Graph (possibly cyclic)
• "chain": LWF-chain graphs
• Additional experimental modes

To reproduce the results from the paper, please follow the following instructions:

1. Clone this GitHub repository

2. Download Download and install clingo to ./inst/asp/, unzip the code from [4] to run the ASP solver there.

3. Install the dagma and CausalDisco Python libraries in a conda environment called `"glip"``.

4. Install the glip package and GUROBI following the instructions above.

To reproduce the timing comparison against ASP, run ./inst/code/run-asp-comparison.R.

To reproduce the simulation study, run ./inst/code/run-simulation.R.

To reproduce the comparison on benchmark datasets, run ./inst/code/run-benchmark-datasets.R.

To reproduce the chain graph simulation, run ./inst/code/run-chain-graph-simulation.R.

The above experiments are time consuming (about 7 days CPU time on a high performance cluster), which is why the scripts above contain default values that make computation faster, but correspond only to a small fraction of experiments that were actually run. The shell scripts in ./inst/slurm/ reproduce all results. Code to produce the figures and tables follows the same naming conventions as above, replacing run with vis.

[1] Kook, L., Mogensen, S. W. Exact Graph Learning via Integer Programming. arXiv preprint arXiv:2601.20589, 2026. doi:10.48550/arXiv.2601.20589

[2] Gurobi Documentation

[3] Kook, L. & Lundborg A. R. Algorithm-agnostic significance testing in supervised learning with multimodal data. Briefings in Bioinformatics 25(6) 2024. doi:10.1093/bib/bbae475

[4] A. Hyttinen, F. Eberhardt, and M. Järvisalo. Constraint-based causal discovery: Conflict resolution with answer set programming. In Proceedings of the 30th Conference on Uncertainty in Artificial Intelligence (UAI), 2014.

• Error: “Solver gurobi not available.” → Ensure GUROBI and the gurobi R package are installed and licensed on your system.

• Parallel execution Some functions accept parallel = TRUE and ncores arguments for faster computation.

• Raise issues or suggest features: Please use GitHub Issues.