Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Learning Mean Field Control on Sparse Graphs
Authors: Christian Fabian, Kai Cui, Heinz Koeppl
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Besides a theoretical analysis, we design scalable learning algorithms which apply to the challenging class of graph sequences with finite first moment. We compare our model and algorithms for various examples on synthetic and real world networks with mean field algorithms based on Lp graphons and graphexes. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering and Information Technology, Technische Universit at Darmstadt, Germany 2Hessian Center for Artificial Intelligence (hessian.AI). |
| Pseudocode | Yes | Algorithm 1 LWMFC Policy Gradient Algorithm 2 LWMFMARL Policy Gradient |
| Open Source Code | No | The paper mentions using third-party libraries (MARLlib, RLlib, PPO) but does not provide any explicit statement or link for the open-sourcing of the authors' own LWMFC or LWMFMARL implementation. |
| Open Datasets | Yes | We use eight datasets from the KONECT database (Kunegis, 2013), where we substitute directed or weighted edges by simple undirected edges: CAIDA (Leskovec et al., 2007)(N 26k), Cities (Kunegis, 2013) (N 14k), Digg Friends (Hogg & Lerman, 2012) (N 280k), Enron (Klimt & Yang, 2004) (N 87k), Flixster (Zafarani & Liu, 2009) (N 2.5mm), Slashdot (G omez et al., 2008) (N 50k), Yahoo (Kunegis, 2013) (N 653k), and You Tube (Mislove, 2009) (N 3.2mm). |
| Dataset Splits | No | The paper does not explicitly provide details about training, validation, or test splits for any of the datasets used. It mentions using 'synthetic CL graphs of size N' and 'empirical datasets' but without split information. |
| Hardware Specification | No | For our experiments, we used around 80 000 core hours on CPUs, and each training run usually took a single day of training on up to 96 parallel CPU cores. The authors also acknowledge the Lichtenberg high performance computing cluster of the TU Darmstadt, but no specific hardware models (e.g., CPU, GPU) are mentioned. |
| Software Dependencies | Yes | We use MARLlib 1.0.0 (Hu et al., 2023a) building on RLlib 1.8.0 (Apache-2.0 license) (Liang et al., 2018) and its PPO implementation (Schulman et al., 2017) for IPPO and our algorithms. |
| Experiment Setup | Yes | For the policies we used two hidden layers of 256 nodes with tanh activations. We used a discount factor of γ = 0.99 with GAE λ = 1.0, and training and minibatch sizes of 4000 and 1000, performing 5 updates per training batch. The KL coefficient and clip parameter were set to 0.2, with a KL target of 0.03. The learning rate was set to 0.00005. |