Learning Mean Field Games on Sparse Graphs: A Hybrid Graphex Approach

Authors: Christian Fabian, Kai Cui, Heinz Koeppl

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate its learning capabilities on both synthetic graphs and real-world networks. This comparison shows that our GXMFG learning algorithm successfully extends MFGs to a highly relevant class of hard, realistic learning problems that are not accurately addressed by current MARL and MFG methods.
Researcher Affiliation Academia Christian Fabian, Kai Cui & Heinz Koeppl Dept. of Electrical Engineering and Information Technology, Technische Universität Darmstadt {christian.fabian, heinz.koeppl}@tu-darmstadt.de
Pseudocode Yes Algorithm 1 Hybrid Online Mirror Descent (HOMD)
Open Source Code Yes The code is available in the supplementary material.
Open Datasets Yes We consider the following real world networks from the KONECT database (Kunegis, 2013): Prosper loans (Redmond & Cunningham, 2013), Dogster (Kunegis, 2013), Pokec (Takac & Zabovsky, 2012), Livemocha (Zafarani & Liu, 2009), Flickr (Mislove et al., 2007), Brightkite (Cho et al., 2011), Facebook (Viswanath et al., 2009), and Hyves (Zafarani & Liu, 2009).
Dataset Splits No The paper does not explicitly provide training/validation/test dataset splits. It mentions simulation on synthetic and real networks, and applying learned equilibria, but no details on how the data was partitioned for training, validation, or testing.
Hardware Specification No The paper states: 'The authors acknowledge the Lichtenberg high performance computing cluster of the TU Darmstadt for providing computational facilities for the calculations of this research.' This identifies a computing cluster but lacks specific hardware details such as CPU/GPU models or memory specifications.
Software Dependencies No The paper does not provide specific software dependencies with version numbers. It mentions 'Python' implicitly through code availability but no other libraries or solvers with versions.
Experiment Setup Yes For the implementation of Algorithm 1, we use an inverse step size (temperature) of γ = 50 and perform 5000 iterations as seen in Figure 2. As degree cutoffs for the periphery, we use kmax = 8 for SIS and SIR, and kmax = 6 for RS... In our experiments, we use τI = 0.2, τR = 0.05, T = 500, µ0(I) = 0.5, c I = 1 and c P = 0.5.