Skip to content
Reinforcement learning using kernel-based function approximation
Branch: master
Clone or download

Latest commit

Latest commit ce0b57e May 22, 2019


Type Name Latest commit message Commit time
Failed to load latest commit information.
algs further cleanup May 10, 2019

Kernel Reinforcement Learning


  • Python 2 or 3
  • OpenAI Gym
  • SciPy
  • MatPlotLib

Available algorithms

  • Kernel Q-Learning with:

    • Continuous states / discrete actions
    • Continuous states and actions from ACC 2018
  • Kernel Normalized Advantage Functions in continuous action spaces from IROS 2018

To run

Kernel Q-Learning with Pendulum with prioritized experience replay

python cfg/kq_pendulum_per.cfg

Kernel NAF with Continuous Mountain Car

python cfg/knaf_mcar.cfg

Other options of configuration files are

  • Kernel Q-Learning for Cont. Mountain Car: cfg/kq_cont_mcar.cfg
  • Kernel Q-Learning for Pendulum: cfg/kq_pendulum.cfg
  • Kernel Q-Learning for discrete-action Cartpole: cfg/kq_cartpole.cfg
  • Kernel NAF for Pendulum: cfg/knaf_pendulum.cfg

Composing policies

The compose folder contains the code for composing two or more trained policies as described in the IROS 2018 paper.

Tuning parameters

To tune learning rates and other parameters, adjust the corresponding parameters in the .cfg file.


This software was created by Ekaterina Tolstaya, Ethan Stump, and Garrett Warnell.

You can’t perform that action at this time.