- Update the complete SAC implementation and modify some details based on the original paper. #365
- Add some extra keyword parameters for
BehaviorCloningPolicyto use it online. #390
- Add functionality for fetching d4rl datasets as an iterable DataSet. Credits: https://arxiv.org/abs/2004.07219
- This supports d4rl and d4rl-pybullet and Google Research DQN atari datasets.
- Uses DataDeps for data dependency management.
- This package also supports RL Unplugged Datasets.
- Support for google-research/deep_ope added.
- Implement
Base.:(==)forSpace. #428
- Add default
Base.:(==)andBase.hashmethod forAbstractEnv. #348
- Add extra two optional keyword arguments (
min_σandmax_σ) inGaussianNetworkto clip the output oflogσ. #428
- Add GaussianNetwork and DuelingNetwork into ReinforcementLearningCore.jl as general components. #370
- Export
WeightedSoftmaxExplorer. #382
- Minor bug & typo fixes
- Removed
ResizeImagepreprocessor to reduce the dependency ofImageTransformations. - Show unicode plot at the end of an experiment in the
TotalRewardPerEpisodehook.
- Make keyword argument
n_actionsinTabularPolicyoptional. #300
- Moved all the experiments into a new package
ReinforcementLearningExperiments.jl. The related dependencies are also removed (BSON.jl,StableRNGs.jl,TensorBoardLogger.jl).
- Add
GraphShortestPathEnv. #445
- Add
StockTradingEnvfrom the paper Deep Reinforcement Learning for Automated Stock Trading: An Ensemble Strategy. This environment is a good testbed for multi-continuous action space algorithms. #428
- Add
SequentialEnvenvironment wrapper to turn a simultaneous environment into a sequential one.
- Set
AcrobotEnvinto lazy loading to reduce the dependency ofOrdinaryDiffEq.