Learning a Prior for Monte Carlo Search by Replaying Solutions to Combinatorial Problems

Fuente: arXiv
Saved in:
Bibliographic Details
Main Author: Cazenave, Tristan
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917570533130240
author Cazenave, Tristan
author_facet Cazenave, Tristan
contents Monte Carlo Search gives excellent results in multiple difficult combinatorial problems. Using a prior to perform non uniform playouts during the search improves a lot the results compared to uniform playouts. Handmade heuristics tailored to the combinatorial problem are often used as priors. We propose a method to automatically compute a prior. It uses statistics on solved problems. It is a simple and general method that incurs no computational cost at playout time and that brings large performance gains. The method is applied to three difficult combinatorial problems: Latin Square Completion, Kakuro, and Inverse RNA Folding.
format Preprint
id arxiv_https___arxiv_org_abs_2401_10431
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Learning a Prior for Monte Carlo Search by Replaying Solutions to Combinatorial Problems
Cazenave, Tristan
Artificial Intelligence
Monte Carlo Search gives excellent results in multiple difficult combinatorial problems. Using a prior to perform non uniform playouts during the search improves a lot the results compared to uniform playouts. Handmade heuristics tailored to the combinatorial problem are often used as priors. We propose a method to automatically compute a prior. It uses statistics on solved problems. It is a simple and general method that incurs no computational cost at playout time and that brings large performance gains. The method is applied to three difficult combinatorial problems: Latin Square Completion, Kakuro, and Inverse RNA Folding.
title Learning a Prior for Monte Carlo Search by Replaying Solutions to Combinatorial Problems
topic Artificial Intelligence
url https://arxiv.org/abs/2401.10431