In this paper, the authors demonstrate a feasible and promising approach to automatically constructing machine learning pipelines.
We formulate the AutoML problem of pipeline synthesis as a single-player game, in which the player starts from an empty pipeline, and in each step is allowed to perform edit operations to add, remove, or replace pipeline components according to a pipeline grammar. A sequence of game steps results in a complete pipeline which is executed on the dataset to solve the task, evaluated by pipeline performance. Formally, an entire pipeline is a state, an action corresponds to modifying the current pipeline to derive a new pipeline, and pipeline performance is the reward. Thus, our approach is based on reinforcement learning.
Additionally,
we use a neural network for predicting pipeline performance and action probabilities along with a Monte-Carlo tree search (MCTS) which makes strong decisions based on the network
Code can be found at AUTOML2019 submission.
Below is the abstract of Automatic Machine Learning by Pipeline Synthesis using Model-Based Reinforcement Learning and a Grammar.
Automatic machine learning is an important problem in the forefront of machine learning. The strongest AutoML systems are based on neural networks, evolutionary algorithms, and Bayesian optimization. Recently AlphaD3M reached state-of-the-art results with an order of magnitude speedup using reinforcement learning with self-play. In this work we extend AlphaD3M by using a pipeline grammar and a pre-trained model which generalizes from many different datasets and similar tasks. Our results demonstrate improved performance compared with our earlier work and existing methods on AutoML benchmark datasets for classification and regression tasks. In the spirit of reproducible research we make our data, models, and code publicly available.
