Skip to content

Implement ddpg

Nicola Gatto requested to merge implement-ddpg into master

General updates:

Changes to CNNTrain2Gluon:

  • Add generation method where an architecture symbol can be passed to the train generator
  • Require an architecture symbol when the training configuration is a reinforcement learning configuration
  • If configuration symbol is a reinforcement configuration with DDPG algorithm, the train generator has to generate a critic network based on a CNN architecture definition file. The class CriticNetworkGenerator is added. This class provides an interface to generate such a critic network.
  • Rework reinforcement learning templates (refactoring, improvements in logging and reusability, etc.)
  • DDPG Agent class is added to Agent.ftl template
  • Add ArchitectureAdapter that is a concrete implementation of the TrainedArchitecture interface defined by CNNTrain. It allows a more convenient access to the architecture symbol.
  • Replay memory supports multi-dimensional actions

Changes to CNNArch2Gluon:

  • Add methods that allow the generation of CNN architectures with multiple inputs (only for python template files)
  • Support multiple architecture port inputs in python templates CNNCreator and CNNNet
  • Add possibility to generate only python files or only cpp files

Merge request reports