Résumé |
Symbolic regression, the task of predicting the mathematical expression of a
function from the observation of its values, is a difficult task which has until
now mainly been tackled with genetic algorithms. The latter involve costly
searches through vast function spaces, and do not leverage past experience: each
new problem is recomputed from scratch.
In the first part of this talk, I will present our recent attempt at solving this
problem via machine learning, by training Transformer models (originally built for
machine translation) on huge datasets of synthetic examples.
In the second part, I will present a specific application: that of recurrence
prediction, i.e. recognising the recurrence relation of number sequences, for
example 1,2,3,5,8->x_n=x_n-1 + x_n-2. |