Optimization of time-varying systems

McBride, L.; Narendra, Kumpati S.

doi:10.1109/tac.1965.1098164

Cited by 14 publications

(3 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…First of all, since the basis of the algorithms described here is the commonly used method of estimating parameters by moving along the gradient of an error measure, it should not be surprising that at least some of these techniques are rediscoveries or simple nonlinear extensions of techniques already described in the engineering literature. In particular, the real-time recurrent learning algorithm coincides with an approach suggested in the system identi cation literature (McBride & Narendra, 1965) for tuning the parameters of an arbitrary dynamical system. Also, the teacher forcing idea appears in the adaptive signal processing literature (Widrow & Stearns, 1985, pp.…”

Section: Relationship To Standard Engineering Approachesmentioning

confidence: 99%

Adaptive State Representation and Estimation Using Recurrent Connectionist Networks

Williams¹

1991

Neural Networks for Control

View full text Add to dashboard Cite

Section: Relationship To Standard Engineering Approachesmentioning

confidence: 99%

Adaptive State Representation and Estimation Using Recurrent Connectionist Networks

Williams¹

1991

Neural Networks for Control

View full text Add to dashboard Cite

“…The real-time recurrent learning (RTRL) utilizes the generality of the backpropagation through time (BPTT) while not suffering from its growing memory requirements in arbitrarily long training sequences. It is similar to the approach proposed by [McBride and Narendra, 1965] for tuning the parameters of general dynamic systems. The RTRL algorithm described in the following follows from [Williams and Zipser, 1989] with unrestricted architectures (see [Haykin, 1999], [Pearlmutter, 1990] and [Pearmutter, 1995] for more details).…”

Section: Teacher -Forced Real-time Recurrent Learning (Rtrl) Algorithmmentioning

confidence: 98%

Development of new model-based adaptive predictive control algorithms and their implementation on real-time embedded systems

Akpan¹

View full text Add to dashboard Cite

This dissertation deals with the development of two new neural network-based model identification algorithms and two new model predictive control (MPC) algorithms which are combined to form model-based adaptive control strategies. Also, two new computer platforms for the implementation of these algorithms and their corresponding strategies are proposed. The overall strategies consist of an online model identification part and a model-based predictive control part. The proposed algorithms for the model identification are practically two new algorithms for training a dynamic feedforward neural network (DFNN) which will be considered to comprise the model of a nonlinear dynamic system. The proposed training algorithms are: the adaptive recursive least squares (ARLS) algorithm and the modified Levenberg-Marquardt algorithm (MLMA). The proposed algorithms for the predictive control parts are: the adaptive generalized predictive control (AGPC) and the nonlinear adaptive model predictive control (NAMPC). The two real-time platforms for the implementation of the combined operation of identification and predictive control algorithms with the purpose of forming an adaptive control strategy are: a service-oriented architecture (SOA) cluster network based on the device profile for web services (DPWS) and a Xilinx's Virtex-5 FPGA (field programmable gate array) development board.The proposed control strategies have been applied to control three applications, namely: the fluidized bed furnace reactor (FBFR) of the steam deactivation unit (SDU) used for preparing catalyst for evaluation in a fluid catalytic cracking (FCC) pilot plant; an activated sludge wastewater treatment plant (AS-WWTP) in accordance with the European wastewater treatment standards; and the auto-pilot control unit of a nonlinear F-16 fighter aircraft. The neural network models for these three applications were validated using one-step, five-step and tenstep ahead prediction simulations as well as with the Akaike's final prediction error (AFPE) estimate algorithm.Also, the performances of the proposed ARLS and MLMA algorithms were compared with the backprogation with momentum (BPM) and incremental backpropagation (INCBP) algorithms. Also the performances of the PID control of the identified model of the FBFR process by means of the ARLS and the MLMA network training algorithms versus the PID control of the first principles model of the same process.The AGPC and NAMPC control of the considered applications when model identification is performed by the ARLS and MLMA algorithms were implemented on a conventional mulitcore personal computer (PC) and SOA cluster of muticore PCs. Hardware-in-the-loop simulations have been performed by linking the PC or SOA implementations with MATLAB simulations of the processes. An AGPC implementation with neural networks trained using the MLMA algorithm has been made on a Xilinx Virtex-5 FPGA.The hardware-in-the-loop simulations have shown that the proposed algorithms and their SOA or FPGA implementations can have execution time...

show abstract

“…Perhaps the earliest reference to it is McBride and Narendra (1965). More recently, it has been independently rediscovered in the context of recurrent neural networks numerous times.'…”

Section: W(k + 1) = W ( K ) + a D ( K ) E ( K ) (11)mentioning

confidence: 99%

Training recurrent networks using the extended Kalman filter

Williams

[Proceedings 1992] IJCNN International Joint Conference on Neural Networks

View full text Add to dashboard Cite

The extended Kalman filter (EKF) can be used as an on-line algorithm to determine the weights in a recurrent network given target outputs as it runs. This paper notes some relationships between the EKF as applied to recurrent net learning and some simpler techniques that are more widely used. In particular, making certain simplifications to the EKF gives rise to an algorithm essentially identical to the real-time recurrent learning (RTRL) algorithm. Since the EKF involves adjusting unit activity in the network, it also provides a principled generalization of the teacher forcing technique. Prelinary simulation experiments on simple finite-state Boolean tasks indicate that the EKF can provide substantial speed-up in number of time steps required for training on such problems when compared with simpler on-line gradient algorithms. The computational requirements of the EKF are steep, but turn out to scale with network size at the same rate as RTRL. These observations are intended to provide insights that may lead to recurrent net training techniques that allow better control over the tradeoff between computational cost and convergence time.

show abstract

Optimization of time-varying systems

Cited by 14 publications

References 7 publications

Adaptive State Representation and Estimation Using Recurrent Connectionist Networks

Adaptive State Representation and Estimation Using Recurrent Connectionist Networks

Development of new model-based adaptive predictive control algorithms and their implementation on real-time embedded systems

Training recurrent networks using the extended Kalman filter

Contact Info

Product

Resources

About