Robust Adaptive Kolmogorov-Arnold Neural Control
Adilkhan Salkimbayev
Overview
The project addresses a gap present in existing control architectures: while Model Predictive Control represents a rigorous paradigm capable of robust and optimal control, it is also computationally demanding and produces a control law that is only piecewise continuous (Bemporad, 2003). The architecture I designed addresses both limitations: by designing the control law to be Lipschitz, that is differentiable a.e. (Rademacher’s Theorem), I managed to reduce the Total Variation of Input by upwards of 1.5x; additionally, the architecture only performs linear matrix operations (Recursive Least Squares with a forgetting factor), which provided me with a runtime considerably reduced compared to the competing controllers.
My role
The project was designed and architected by me from start to finish.
Approach
Existing literature was used in combination with findings from my first research project. Most of the techniques were implemented and coded by hand in the IDE; the techniques featured has been simulated in Python and saved in corresponding CSV files for further parsing and presentation.