Neural Architecture Search poses a problem for Deep Learning. We use a Reinforcement Learning method to solve this issue, with multiple worker agents. The approach uses multiple DDPG agents, which are run by a controller and treats the environment of the dataset as a input to work on.
- Model: variable number of nodes, fixed number of hidden layers
Framework: PyTorch
Mode of Update: Asynchronous
Algorithm: Multi-Agent DDPG
Environment: Datasets
Neural Network: DNNs
Sampling: Replay Buffer
Type of Learning: Actor-Critic Method
The controller can be trained on GPUs as well.
torch
numpy
pandas
sklearn
git clone https://github.com/wolflegend99/nas-rl.git