Shrinking Window Optimization Algorithm Applied to Pneumatic Artificial Muscle Position Control

Pneumatic Artificial Muscles (PAMs) are compliant actuators that can be safely used to interact with humans. Additionally, they can be made compatible with functional Magnetic Resonance Imaging (fMRI) for neurorehabilitation procedures and MRI-guided surgery. A major drawback of PAMs, however, is the precise positioning control due to its highly nonlinear dynamics. This control problem can be solved using an optimal control approach, defining an objective function and solving for the controller’s parameters that minimizes this function. Typically this is done using a model of the system, and further trial and error tuning is required for better performance in the implementation stage, to compensate for the imperfections of the model. Optimizing the parameters using the real system to evaluate the objective function is harder because of the time, physical limitations and noise of the objective function, but it can provide better quality parameters and eliminate further tuning. For that, we use a novel optimization algorithm called Shrinking Window (SW), which is capable of finding the controller’s parameters faster and with similar quality to state-of-the-art global optimization algorithms such as Bayesian Optimization (BO). We describe the SW algorithm and compare its performance on a synthetic noisy function and in tuning the controller for a positioning system powered by a PAM, against random search and BO. On learning the controller’s parameters, we report a 48,15% shorter learning time and a 8% improvement on parameter quality compared to BO.