Learning a Prior for Monte Carlo Search by Replaying Solutions to Combinatorial Problems
Fuente:
arXiv
Saved in:
| Main Author: | |
|---|---|
| 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 |