Skip to content

Latest commit

 

History

History
102 lines (70 loc) · 3 KB

NEWS.md

File metadata and controls

102 lines (70 loc) · 3 KB
Raw

Release History

Version 1.1

Not released yet.

Features

  • New representation: ProMPBehavior
  • New environment: OptimumTrajectoryCurbingObstacles

Bugfixes

  • DMPBehavior: Use pseudo inverse in imitation learning for DMPs. Imitating trajectories with less points than DMP weights is now possible.
  • DMPBehavior: Fix issue with many decimal places in dt.
  • Optimizer: Is now an abstract base class as it should have been.
  • REPSOptimizer / CREPSOptimizer: Use analytical gradient for optimization. This improves the computational efficiency and gives slightly better results.

Version 1.0

2018/02/11

Maintenance

  • Make BOLeRo compatible to latest version of OpenAI Gym (0.9.6)

Breaking Changes

  • C++: BehaviorSearch, Optimizer, and Environment have an additional parameter 'config' in their 'init' function. The YAML-based configuration string will be given. Each implementation of this function must be able to cope with an empty config string. In this case all parameters will be set to their default values.
  • C++: The controller will run 'reset' of the environment once directly after 'init' before the first episode is executed.
  • C++: Controller only passes corresonding sections of the learning configuration to the components.
  • pso_optimizer: uses config section 'Optimizer' (previously: 'BehaviorSearch Parameters').

Version 0.2

2018/01/23

Features

  • New ContextualOptimizer: C-CMA-ES (based on CMA-ES)
  • Support for Windows and MacOS

Documentation

  • Documented context_features of CREPSOptimizer

Version 0.1

2017/12/19

Features

  • Continuous integration with Travis CI and CircleCI
  • Added docker image
  • New behavior search: Monte Carlo RL
  • New optimizer: relative entropy policy search (REPS)
  • New optimizer: ACM-ES (CMA-ES with surrogate model)

Bugfixes

  • DMPSequence works with multiple dimensions
  • Minor fixes in docstrings
  • Multiple minor fixes for Travis CI
  • Fixed scaling issues in C-REPS

Documentation

  • Documented merge policy
  • Added meta information about the project to the manifest.xml
  • Updated documentation on how to build custom MARS environments

Version 0.0.1

2017/05/19

First public release.

Breaking Changes

In comparison to the old behavior learning framework used by the DFKI RIC and the University of Bremen, we changed the following details:

  • Python interface: changed signature int Environment.get_feedback(np.ndarray) to np.ndarray Environment.get_feedback()
  • Python interface: ContextualEnvironment is now a subclass of Environment
  • Python interface: renamed Environment.get_maximal_feedback to Environment.get_maximum_feedback
  • Python and C++ interface: Behavior constructor does not take any arguments, instead the function Behavior.init(int, int) has been introduced to determine the number of inputs and outputs and initialize the behavior
  • Python interface: Optimizer and ContextualOptimizer are independent
  • Python interface: BehaviorSearch and ContextualBehaviorSearch are independent