• Search by Module
• Search by Words
• Search Projects

• Java
• Python
• JavaScript
• TypeScript
• C++
• Scala
• Blog

Project: kusanagi (GitHub Link)

• kusanagi-master
  • .hgignore
  • Doxyfile
  • examples
    • run_learning_algorithm.py
    • transfer
      • cartpole_adjust.py
    • notebooks
      • evaluate_policy-Copy2.ipynb
      • gym_learn.ipynb
      • arduino_cartpole_environment.ipynb
      • kl_loss
        • task_cost_plus_kl_reg.ipynb
      • evaluate_policy.ipynb
      • plotting_test.ipynb
      • torch_experiments.ipynb
      • cartpole_learn.ipynb
    • data
    • cartpole_learn.ipynb
    • __init__.py
    • PILCO
      • double_cartpole_learn.py
      • cartpole_learn.py
      • cartpole_learn_serial.py
      • __init__.py
      • pendulum_learn.py
  • kusanagi
    • utils
      • probls.py
      • utils_.py
      • distributions.py
      • updates.py
      • __init__.py
      • extra_ops.py
    • shell
      • double_cartpole.py
      • plant.py
      • pendulum.py
      • arduino
        • cartpole
          • cartpole.ino
        • arduino.py
        • __init__.py
      • cost.py
      • evaluate_policy.py
      • double_integrator.py
      • __init__.py
      • experiment_utils.py
      • cartpole.py
    • server.py
    • ghost
      • transfer
        • trajectory_matching.py
        • __init__.py
        • kl_matching.py
      • algorithms
        • pilco.py
        • pddp.py
        • mc_pilco.py
        • __init__.py
      • regression
        • SPGP.py
        • SNRpenalty.py
        • objectives.py
        • BaseRegressor.py
        • GP.py
        • NN.py
        • nonlinearities.py
        • cov.py
        • layers.py
        • SSGP.py
        • __init__.py
      • __init__.py
      • optimizers
        • sgd_optimizer.py
        • __init__.py
        • scipy_optimizer.py
      • control
        • control_.py
        • NNPolicy.py
        • saturation.py
        • __init__.py
    • base
      • base_.py
      • ExperienceDataset.py
      • Loadable.py
      • __init__.py
    • diagnosis_tools
      • profile_learner_log.py
      • profile_learner_logs.sh
      • check_init_exp_aqua.py
      • __init__.py
      • plot_learner_profile.m
    • __init__.py
  • LICENSE
  • test
    • test_serial_plant.py
    • testNNRegression.py
    • matlab_test.m
    • __init__.py
    • testGPRegressor.py
  • cartpole_experiments.sh
  • setup.py
  • eval_cartpole.sh
  • README.md
  • double_cartpole_experiments.sh
  • .gitignore

kusanagi

Recommended way to install:

Install the Miniconda 3 distribution: https://conda.io/miniconda.html

conda install numpy scipy mkl mkl-rt mkl-service jupyter qt
conda install pygpu libgpuarray -c mila-udem
pip install --upgrade git+https://github.com/Theano/Theano
pip install --upgrade git+https://github.com/Lasagne/Lasagne
pip install gym
pip install mujoco-py
cd <KUSANAGI_ROOT>
pip install -e .

If you use the Miniconda 2 distribution (python 2.7.x), then you need to install gym 0.5.7 pip install gym==0.5.7

Example to reproduce some of the results:

python examples/PILCO/cartpole_learn.py -e 8 -n mcpilco_lognormal_dropout -k mc_samples 100 -k max_evals 1000 -k learning_rate 1e-3 -k polyak_averaging None -r True -k clip_gradients 1.0

Results on robot learning with ROS

See http://github.com/juancamilog/robot_learning for an example on how to integrate this library with ROS, and some sample results (from our IROS 2018 submission "Synthesizing Neural Network Controllers with Probabilistic Model based Reinforcement Learning")