Master Thesis by Erik Wannerberg

This repository containts the source code used in my Master's thesis, Automatic Heuristic Generation for Optimal Control. Feel free to browse, download, test, and modify at your heart's desire.

The code is licensed under the MIT License, which allows precisely what is suggested above, with the only requirement that you take the license file with you if you ship it somewhere else.

The thesis itself can be found on the GitHub wiki (direct link to thesis here).

The code is split up into four parts:

The ModelPredictiveControl folder contains a python framework for nonlinear Model Predictive Control (MPC) using the python optimization framework pyomo. It has the following dependencies:

• python3
• pyomo itself, provided through pip
• numpy
• scipy
• Ipopt (optimization solver). Source code and binaries are available at their homepage.

The MPC implementation with a model of a car on a hilly road is run by invoking python with run.py followed by the path to a JSON configuration file, of which an example is available at example/test.json. The path to the ipopt solver binary is specified there.

The DiffusionMaps&ClosedObservables folder contains mainly MATLAB code for nonlinear dimensionality reduction using diffusion maps. The code is heavily inspired from Felix Dietrich's approach of closed observables, and attempts to produce an accurate model of the dynamics of the controlled MPC variables.

It of course requires MATLAB, and uses the MATLAB Statistics and Control Toolbox for a convenient pairwise distance function. It is possible to encode this function in a few lines, so feel free to do so if you do not have this toolbox available.

The few lines of python code in there for data pruning have the same requirements as above, except for pyomo and ipopt.

A diffusion maps example can be run by running the dmapsTesting script, which uses an output file generated from a scenario with a height profile generated by 3 sinusoidal curves different amplitudes and frequencies.

The Validation folder contains MATLAB code for validating results of the previous frameworks, and relies on the function in runScenario.m to run a simulation for a user-specified scenario with a user-provided controller, following the specifications of those provided in the folder. run_dmaps_testing.m contains an example script to construct a controller and run it in the framework. You can try it without modification, but it probably won't work, as the controller relies on varied data to produce sensible results.

For acquiring real height data, the HeightDataRetrieval contains python functions for downloading such using the Google Maps Elevation API as well as the Directions API. To use this, you need a Google Maps API key, which is available for free, linked at the API documentation. The key should be entered in get_key.py, where it should be returned from the get_key function, instead of raising an exception.

The code depends on the googlemaps module in python, which again is available through pip.

Any questions or requests for the master thesis are happily accepted. Have fun using!