Neurocontroller alternatives for "fuzzy" ball-and-beam systems with nonuniform nonlinear friction

Area of Science:

• Control Systems Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• The ball-and-beam problem is a standard benchmark for control algorithm evaluation.
• Professor L. Zadeh proposed a modified version requiring a fuzzy logic controller, involving a beam with a sticky substance.
• Predicting ball motion is further complicated by omitting velocity information.

Purpose of the Study:

• To develop and test a novel neurocontroller for a modified ball-and-beam problem.
• To investigate the use of recurrent neural networks without explicit velocity inputs.
• To explore the application of adaptive critic designs, specifically dual heuristic programming (DHP), for physical system control.

Main Methods:

• Recurrent neural networks were employed, using only consecutive positions as input.
• Truncated backpropagation through time with the node-decoupled extended Kalman filter (NDEKF) algorithm updated network weights.
• An adaptive critic design, specifically dual heuristic programming (DHP), was utilized as the neurocontroller architecture.

Main Results:

• The developed neurocontroller successfully controlled the modified ball-and-beam system.
• This represents the first known application of DHP to control a physical system.
• The system responded to Professor Zadeh's challenge, demonstrating a novel approach.

Conclusions:

• The study demonstrates the feasibility of using recurrent neural networks and DHP for complex control tasks without velocity inputs.
• This work contributes to the ongoing scientific dialogue regarding the synergy between neural network control and fuzzy logic systems.
• While not claiming superiority, the neurocontroller offers a valuable contribution to the field.