Compositional Optimization

Abstract

This talk presents ongoing work with David Spivak on compositional learning and optimization. It is an attempt to unify the compositional theory of dynamical systems in the category of polynomial functors with concepts from optimization, multi-agent reinforcement learning and information theory. A preliminary framework is presented that allows to talk about the composition of certain classes of optimizing systems. If time permits, an example from reinforcement learning or information theory may be presented at the end.