# Automated Machine Learning with Monte-Carlo Tree Search (Extended Version)

**Authors:** Herilalaina Rakotoarison, Marc Schoenauer, Michèle Sebag  
**arXiv:** [1906.00170](https://arxiv.org/abs/1906.00170)  
**Date:** 2019-06-01  
**Summary page:** https://www.summarizepaper.com/en/paper/1906.00170v1/

## Key points

- Automated Machine Learning (AutoML) involves selecting the most suitable algorithm and determining hyperparameter values for optimal performance on a dataset.
- Mosaic is a Monte-Carlo tree search (MCTS) based approach designed to address complex AutoML problems by combining structural and parametric optimization in an expensive black-box setting.
- The study includes empirical investigations comparing optimization processes based on Bayesian optimization versus MCTS, exploring warm-start initialization techniques, and assessing the benefits of ensembling solutions gathered during the search process.
- Mosaic outperformed Auto-Sklearn in tests on both OpenML 100 benchmark dataset and Scikit-learn portfolio, showing statistically significant improvements in AutoML tasks by leveraging innovative MCTS-based strategies.

## Summary

In the field of Automated Machine Learning (AutoML), the task involves selecting the most suitable algorithm from a machine learning portfolio and determining its hyperparameter values to achieve optimal performance on a given dataset. In this extended version, authors Herilalaina Rakotoarison, Marc Schoenauer, and Michèle Sebag introduce Mosaic, a Monte-Carlo tree search (MCTS) based approach designed to address the complex AutoML problem that combines structural and parametric optimization in an expensive black-box setting. The study includes extensive empirical investigations to independently evaluate and compare various aspects of the optimization process. This includes comparing the effectiveness of optimization processes based on Bayesian optimization versus MCTS, exploring the impact of warm-start initialization techniques, and assessing the benefits of ensembling solutions gathered throughout the search process. Mosaic is put to the test on both the OpenML 100 benchmark dataset and the Scikit-learn portfolio. The results reveal statistically significant improvements over Auto-Sklearn, which had previously emerged as a top performer in international AutoML challenges. The findings underscore Mosaic's prowess in delivering enhanced performance in AutoML tasks by leveraging innovative MCTS-based strategies for algorithm selection and hyperparameter tuning.