-------------------------------------------------
SC42050 Knowledge-based control systems
-------------------------------------------------
REPRODUCIBILITY ASSIGNMENT
- A Reinforcement Learning-Based Architecture for Fuzzy Logic Control
The code related to the reproducibility assignment is collected in the Python
files in this repository.
System requirements: Python 3.x (might be compatible with Python 2.7, but it's not tested)
Dependencies:
- gym
- numpy
- scipy
- matplotlib
The following executables are related to the controllers
and tests performed on them:
- run_ARIC.py : Code related to the training of the ARIC controller
NOTE: ARIC learning is unstable. Sometimes works and sometimes the random initial weights lead
to dead-end solutions where the learning gets stuck, as described in the report. If the
controller does not learn in the first 10 attempts, just re-run the file.
- run_LQR.py : Code related to the testing of the LQR controller
- run_Q_agent.py : Code related to the training of the Q-learning controller
- run_sensitivity : Code related to the sensitivity test performed on the ARIC
The following files contain functionality imported by the executables above:
- ARIC_model.py : The actual implementation of the ARIC controller as described by Berenji
- membership_functions.py : Definitions of the membership functions for ARIC inputs and outputs
- fuzzy_set.py : Handy class for representing and working with fuzzy functions
- pole_cart.py : Own implementation of the pole-cart model
- utils.py : Small handy functions that reduce clutter code in the implementations