This is my bachelor's thesis project. It is an implementation of the DDPG algorithm in a classic, custom cart-pole environment created by my mentor Dr. Domen Šoberl. Some pre-trained weights are to be found here, which can be used. Important to note, there are some "flags" to be found in the "util" module, with which certain logs can be controled (also the ones added subsequently, if necessary), as well as rendering and recording.
This is not a comprehensive list of all the packages used in this project, however, the packages listed here are the minimum required dependencies to run the project. Additional dependencies will be necessary (e.g. pyplot etc.) for plotting data gathered from training.
Use `conda` (miniconda3) to install TensorFlow and all its dependencies - as per the official instructions found on TensorFlow website.
Use `pip` to install `tf-agents` WITHIN the conda environment where TensorFlow is set up.
Use `pip` to install `gymnasium` - can be a global install (not specific to the conda environment).
Use `pip` to install `tkinter` - can be a global install (not specific to the conda environment).
Use `pip` to install `pygame` - can be a global install (not specific to the conda environment).
Use `pip` to install `numpy` - can be a global install (not specific to the conda environment).
The project is structured so that its components are grouped into modules. The main module is the "ddpg", where the DDPG class and the main implementation of the algorithm in the custom Cart-Pole environment are stored, in "ddpg.py" and "cartpole.py" respectively. Therefore, the recommended way of running the main cartpole.py is:
python3 -m ddpg.cartpole
(In my case, for Debian based distro, the "python3" command is used - otherwise for RPM, Arch based distros use "python")
Similarly, running any other file from the project follows the same scheme e.g. python3 -m example.pendulum.
For gathering plotting data, there are two simple shell scripts that can be used.