Reproduce “Model-free policy gradient for discrete-time mean-field control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Matthieu Meunier, Huyên Pham, Christoph Reisinger
The authors introduce a policy-gradient method for finite-state, finite-horizon mean-field control. It perturbs logits of the population distribution and combines trajectories with an estimate of its sensitivity. Under assumptions, the perturbed gradient converges to the true gradient as perturbation shrinks. In a two-state, two-action test, policy errors were lower at ε=0.2 than ε=2.0; this within-method result is not against another algorithm.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.