First-Order Methods for Large-Scale Market Equilibrium Computation

Authors: Yuan Gao, Christian Kroer

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments show that Proportional Response dynamics is highly efficient for computing approximate market equilibria, while projected gradient with linesearch can be much faster when higher-accuracy solutions are needed. We perform numerical experiments on market instances under all three utilities with various generated parameters.
Researcher Affiliation Academia Yuan Gao Department of IEOR, Columbia University New York, NY, 10027 gao.yuan@columbia.edu Christian Kroer Department of IEOR, Columbia University New York, NY, 10027 christian.kroer@columbia.edu
Pseudocode Yes We prove that proximal gradient with a non-standard practical linesearch scheme (Algorithm 1 in the appendix) converges linearly
Open Source Code Yes Codes for the numerical experiments are available at https://github.com/Coffee And Convexity/fom-for-me-codes.
Open Datasets No For linear utilities, we generate market data v = (vij) where vij are i.i.d. from standard Gaussian, uniform, exponential, or lognormal distribution. For each of the sizes n = 50, 100, 150, 200 (on the horizontal axis) and m = 2n, we generate 30 instances with unit budgets Bi = 1 and random budgets Bi = 0.5 + Bi (where Bi follows the same distribution as vij).
Dataset Splits No The paper describes generating synthetic data and uses a termination criterion based on convergence error, not standard training/validation/test dataset splits.
Hardware Specification No Part of the numerical experiments were run on the computing server of Columbia University Data Science Institute https://datascience.columbia.edu/about-us/work-with-us/computing-resources/. This only states the location, not specific hardware components like CPU/GPU models.
Software Dependencies No The paper mentions CVXPY and Mosek but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes For PGLS, we report the number of linesearch iterations (that is, the total number of projection computations). For other algorithms, we report the number of iterations. As a fair comparison, we use the same parameters α, β, Γ for PGLS throughout without handpicking.