NadeemWard / kernel-based_RL

Reimplementation of "Kernel-Based Reinforcement Learning" Ormoneit and Sen (2002) (https://link.springer.com/content/pdf/10.1023/A:1017928328829.pdf)

Kernel-Based Linear Program

This codebase provides an implementation of a Kernel-Based Linear Program for solving MDPs. This work is based on "Kernel-Based Reinforcement Learning" Ormoneit and Sen (2002) (https://link.springer.com/content/pdf/10.1023/A:1017928328829.pdf) and also provides an implementation of their approximate value iteration algorithm.

Requirements

• CVXPY
• sklearn
• gym

Running experiments.

Example.ipynb is an example jupyter notebook demonstrating functionality.

Sample experimental results.

Both the Kernel-Based Linear Program and Kernel-based value iteration algorithms where run on OpenAI's Cartpole environment. This environment has a continuous state space and 2 discrete actions (left and right). It is considered solved when the reward reaches 200.

Solving CartPole using a Kernel-Based LP

Solving CartPole using Kernel-Based Value Iteration