Scripts to test various reinforcement learning algorithms on simple control problems.
This is a work-in-progress...
To start off, I developed a new class that implements a linear quadratic regulator (LQR). This is a class of controllers used in control engineering applications to stabilize dynamical systems.
To find the optimal controller for a given dynamical system you ideally need to know the system dynamics. The following module uses the Python Control Systems Library to find an LQR for a cart-pendulum system:
Now, I'm trying out some of the Stable Baselines implementations of reinforcement learning (RL) algorithms.
I'm testing these on some new environments I created in OpenAI's Gym toolkit. They are a bit more challenging than the 'classic control' problems on OpenAI's website but east enough to understand.
To check these new environments out, see my other repo:
For a quick demo, run the following script with your choice of environment. The -r option tells it to render the environment (an animation window appears). This will show the environment with no agent controlling it.
python test_run.py -e CartPole-BT-dL-v0 -r
To run an environment with an optimal linear controller, run this script:
python test_run_lqr.py -e CartPole-BT-x2-dL-v0 -r