Project 4: Machine Learning for PID Control

This repository contains the supporting files for our CSCE 635 project submission. This includes:

• MATLAB, Python, and C++ code
• Simulink models
• Simulation data
• Jupyter notebooks for data analysis
• Report in LaTeX

C++ Code - src

The souce code was developed for implementation into EMILY's existing platform. The GPID class is a stand alone class that can be utilized for multiple applications. The GPID class can be simply used by generating an object, and then calling the propagate_net method each loop with error and measurement as inputs, respectively. The GPID class utilizes the gpid_setup.txt file to initialize the starting weights, and toggle learning mode on/off. The gpid_setup file is a TAB delimited file, and is structured as [training_bool P_neuron_weight I_neuron_weight D_neuron_weight].

The Control class is an adaptation from the existing Control class to implement the GPID and LOS control algorithms.

MATLAB and Simulink Code - sim

The simulations were executed with a proof-of-concept implementation of the GPID algorithm in Python, which can be found in the module gradient_tuning.py. The two models sim_GPID_LOS.slx and sim_P_LOS.slx contain the simulation dynamics and must be executed by running exec_sim_main.m.

The simulations require the following MATLAB m-files:

• init.m, which initializes the Python network and trainer objects.
• intertrial.m, which clears the network memory between trials
• structures.m or structures_random.m, which intializes environment variables and parameters for the trials
• update_wrap.m which updates the Python network trainer object and logs the states
• param.mat, contains EMILY's physical parameters

The simulations also require the Simulink library emily.slx, which contains the definitions of the dynamic model and LOS controllers.

Jupyter Notebooks - jupyter

The jupyter notebooks serve to document the analysis and the results in a single framework. The following notebooks are included:

• analysis-dev.ipynb, which develops the tools for data analysis
• analysis.ipynb, which analyzes the trial data
• net-analysis.ipynb, which studies the properties of the network over many training trials
• analysis-final.ipynb, contains comparison of different network parameters and the results of structured trials

LaTeX - report

The report and .tex file and supporting figures.