This paper describes a fast on-line neuro-fuzzy controller for underwater robots of which the dynamics are highly nonlinear, coupled, and time-varying. The neuro-fuzzy controller is based on the Fuzzy Membership Function Neural Network (FMFNN) with varying learning rate. Even though FMFNN has advantages of fuzzy logic and neural networks, it is not so easy to decide initial rules for fuzzy inference part and learning rates for neural network part. These two factors affect controller performance so much. There are many research results about how to decide or modify fuzzy rules. However learning rate was not an issue in neuro-fuzzy controller. To show the validity of varying learning rate method, simulation results of FMFNN with varying learning rate are presented with underwater robot control example.
Autonomous underwater vehicles (AUVs) are well known as a complex dynamic system that is highly nonlinear, coupled, and timevarying (Yuh, 2000). Various advanced underwater robot control system have been developed, such as sliding control (Healey, 1993 and Yoerger, 1993), nonlinear control (Nakamura, 1992), adaptive control (Nie, 1998, and Yuh, 1990b), neural network (Ishiii, 1998, and Yuh, 1990a), fuzzy control (DeBitetto, 1994, and Kato, 1995) and neurofuzzy control (Wang, 2003, and, Kim, 1992). Most controllers are focused on self-tuning or self-organizing when the control performance degrades during the operation due to model uncertainties and changes in the vehicle system and its environment. Among them, on-line neuro-fuzzy controller based on Fuzzy Membership Function Neural Network (FMFNN) showed good performance for nonlinear dynamic system, such as AUVs, without a prior knowledge of nonlinear dynamic system (Kim, 2002). It combined advantages of fuzzy logics and neural network such as adoptability of the human operators experience, inference capability for unknown situations using given knowledge, universal approximation, and learning capability.