Offline reinforcement learning for neural architecture search
A project lead by Martin Cepeda, Romain Egele and Jeremy Gozlan during the reinforcement learning lecture of the AIViC master at Ecole polytechnique.
The repository is composed of an RL environment env/nas_rl.py.
The data_generation.py was used to generate our artificial datasets.
The different notebooks are:
Classic RL Environment: a regular environment for NAS with on-policy and off-policy RL.Offline RL Environment: a environment for NAS with offline RL.Supervised learning on the Dataset: a context-based bandit approach for NAS.Offline DQN: a DQN algorithm adapted to the offline RL setting.