Release Version 2.0

All

  • All internal components now uses TensorL2D instead of MatrixL for full compatibility with DataPredict Neural. TensorL2D can be replaced with TensorL.

  • All constructors now requires a parameter dictionary instead of arguments.

Models

  • Added SoftActorCritic, DeepDeterministicPolicyGradient and TwinDelayedDeepDeterministicPolicyGradient.

  • DeepQLearning, DeepStateActionRewardStateAction, DeepExpectedStateActionRewardStateAction, ProximalPolicyOptimization models and its variants now have “lambda” argument for TD-Lambda and GAE-Lambda functionality. This includes AdvantageActorCritic model.

  • The diagonalGaussianUpdate() function now requires actionNoiseVector.

  • All reinforcement learning models now require “terminalStateValue” for categoricalUpdate(), diagonalGaussianUpdate() and episodeUpdate() functions.

  • Reimplemented ActorCritic, VanillaPolicyGradient and REINFORCE models.

  • Removed AsynchronousAdvantageActorCritic.

AqwamCustomModels

  • Removed “AqwamCustomModels” section and its models.

Others

  • Moved RandomNetworkDistillation, GenerativeAdversarialImitationLearning and WassersteinGenerativeAdversarialImitationLearning to “ReinforcementLearningStrategies” section.

  • Renamed DistributedGradients to DistributedGradientsCoordinator and moved to “DistributedTrainingStrategies” section.

  • Renamed DistributedModelParameters to DistributedModelParametersCoordinator and moved to “DistributedTrainingStrategies” section.

  • Made major changes with ModelChecker, ModelDatasetCreator and OnlineTraining.

QuickSetups

  • Fixed CategoricalPolicyQuickSetup and DiagonalGaussianPolicyQuickSetup, where the next action is used in the categoricalUpdate() and diagonalGaussianUpdate() instead of previous action.

  • CategoricalPolicyQuickSetup and DiagonalGaussianPolicyQuickSetup no longer have setClassesList() and getClassesList() functions.