This project is the implementation of DeepLine, a system for automatic machine learning pipeline generation. DeepLine is a deep reinforcement learning-based framework. It consists of a grid-world environment in which the pipeline generation process is modeled in a semi-constrained setting, and a DDQN agent.
DeepLine also introduces a novel method for handling large discrete action spaces where most of the actions are invalid in most states. The Hierarchical Step filters only the set of vaild actions by quering the agent iteratively on small subsets of the actions space.
Extensive description of the framework is available in deeplineThesis.pdf.
After cloning the repository, do the following:
cd gym-deepline
pip install -e .The code is implemented in Python 3.6.
DeepLine's environment is implemented as an OpenAI-Gym environmnet, and can be called by:
import gym
import gym-deepline
env = gym.make('deepline-v0')The environment itself is implemented in envs/atml_env.py and the three versions of the agent are under agents/.
For a more detailed example, see example.py.