Hyperparameter optimization using agents for large scale machine learning

Abstract

Machine learning (ML) has become an essential tool for humans to get rational predictions in different aspects of their lives. Hyperparameter algorithms are a tool for creating better ML models. The hyperparameter algorithms are an iterative execution of trial sets. Usually, the trials tend to have a different execution time. In this paper we are optimizing the grid and random search with cross-validation from the Dislib [1] an ML library for distributed computing built on top of PyCOMPSs[2] programming model, inspired by the Maggy [3], an open-source framework based on Spark. This optimization will use agents and avoid the trials to wait for each other, achieving a speed-up of over x2.5 compared to the previous implementation.