Learning Graphon Mean Field Games and Approximate Nash Equilibria

Authors: Kai Cui, Heinz Koeppl

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we are able to demonstrate on a number of examples that the finite-agent behavior comes increasingly close to the mean field behavior for our computed equilibria as the graph or system size grows, verifying our theory.
Researcher Affiliation Academia Kai Cui & Heinz Koeppl Department of Electrical Engineering, Technische Universität Darmstadt, Germany {kai.cui,heinz.koeppl}@bcs.tu-darmstadt.de
Pseudocode Yes Algorithm 1 Fixed point iteration Algorithm 2 Backwards induction Algorithm 3 Forward simulation Algorithm 4 Sequential Monte Carlo
Open Source Code Yes For reproducibility, in the supplement we provide all code required to reproduce all results in this work.
Open Datasets No The paper describes the generation of synthetic graphs (
Dataset Splits No The paper does not provide specific dataset split information (e.g., percentages, sample counts for training, validation, or test sets). The experiments are simulations on generated graphs, not based on pre-split datasets.
Hardware Specification No We ran each trial of our experiments on a single conventional CPU core, with typical wall-clock times reaching up to at most a few days. We estimate the required compute to approximately 6500 core hours. We did not use any GPUs or TPUs.
Software Dependencies Yes For PPO, we used the RLlib implementation by Liang et al. (2018) (version 1.2.0, Apache-2.0 license).
Experiment Setup Yes As for the specific configurations used in the PPO experiments, we give the hyperparameters in Table 1 and used with a feedforward neural network policy consisting of two hidden layers with 256 nodes and tanh activations, outputting a softmax policy over all actions.