A Hybrid Subsymbolic System for Controlling Distributed Parameter Systems

摘要

We control a distributed parameter system following a self-tuning model reference adaptive control strategy supported by a recurrent neural network. The key idea is to combine the dynamics of the arm with those of a properly designed recurrent neural network. This lets us train the resulting hybrid system to track a reference trajectory, avoiding the well known problem of explicitly converting tracking errors into control errors. We obtain a system which controls the rotation of a single-link flexible arm carrying a payload along a planar trajectory with two distinguishing features: (ⅰ) the arm reaches the end of the planned trajectory quickly with very smooth dynamics and zero vibration; (ⅱ) the system performs well in a wide range of stop angle-payload mass combinations, where the angle is assigned by the user and the mass is directly identified by the system during rotation.