Séminaire de géométrie - Salle de conférences du CMLS - 10h

10h-11h – Carlos Simpson (Université Côte d'Azur) 

"Machine learning for dynamical programming problems: semigroups, triangulations, going towards proofs"

We discuss a general type of problem that seems to be susceptible to machine learning. In a typical dynamical programming situation the algorithm may require some choices at each stage, and the running time can depend on which choices are made. One would like to minimize the running time or other size measure. This may be set up as a "reinforcement learning" (RL) problem in which the data are self-generated. We discuss two instances: the classification of finite semigroups, and the counting of triangulations of planar lattice polygons. We will comment on how a next possible case, the search for proofs, will enter into the ERC Synergy project "Malinca" that will be starting soon.