Deep deterministic policy gradients (DDPG) is an off-policy learning algorithm for continuous action spaces. It is an extension of deep Q-networks, where the selection of the maximum-value action is approximated using a neural network, rather than computed directly, due to the continuous action space.
scripts/- contains the models, loss functions, and other utilities (e.g. replay buffer) that are used in the DDPG algorithm.
nb/- contains the notebook with the implementation of the DDPG training routine, as well as code for testing an existing model.
- Spinning Up: Detailed explanation and high quality implementation from OpenAI of the DDPG algorithm.