Fully parallel on-chip learning hardware neural network for real-time control


A parallel hardware neural network with on-chip learning ability is presented. The chip is used to perform real-time output feedback control on a nonlinear dynamic system. The non linear plant is a simulated unstable combustion process and is nonlinear enough that linear controllers give poor performance. Neural networks provide an adaptive sub-optimal control that does not need any prior knowledge of the system. The hardware neural network presented here utilizes parallelism to achieve speed independent of the size of the network, enabling real-time control. Parallel on-chip learning ability allows the hardware neural network to learn on-line as the plant is running and the plant parameters are changing. The experimental setup used to show that the parallel hardware learning neural network chip can control the simulated combustion system is described, and the results discussed. 1. NEURAL NETWORK CONTROL Artificial neural networks (ANN) try to emulate the human neural system [1]. They have been experimentally demonstrated to be robust, fault-tolerant, and flexible. In addition they are inherentily parallel and as a result fast when fabricated in approporate hardware. Neural networks can be used to model and control complex nonlinear physical systems with unknown or slowly varying plant parameters [2]. They have been successfully applied to robot arm control, chemical process control, continuous production of high-quality parts, and aerospace applications. A neural network can be a general purpose adaptive controller. In conventional control methods, an extensive study and understanding of the system is required before a controller can be built. And then the controller will only work for the system it was designed to control. Neural networks do not need prior knowledge of a system in order to control it, the same neural network can often be reused for many dissimilar applications. Another advantage of neural networks is potential high speed, when built in hardware. The neural system solves complicated problems by the parallel operation of neurons. Software simulation on serial computers sacrifices this parallelism. However, semiconductor technology provides Our research is part of the Multidisciplinary University Research Initiative (MURI) on Intelligent Turbine Engines (MITE) project, supported by DOD-Army Research Office, under Grant No. DAAH04-96-1-0008. a powerful means to implement artificial neural network with the desirable high degree of parallelism. Currently proofs of stability exist only for special neural network controllers, thus general neural network controllers are not desirable when the plant is well understood and a conventional controller has been proven stable. If the plant is a linear system, a very good controller can be computed from the plant model, and typically will yield much better performance than a neural network. 2. HARDWARE NEURAL NETWORKS FOR REAL-TIME CONTROL There are two general categories for hardware neural networks: digital circuits and analog circuits. Digital circuits have the advantage of implementing precise computational units like adders and shifters, which perform exactly the same functions as computer software. Analog circuits experience non-idealities, like non-linear multipliers, offsets, leakage of weight values stored capacitively, random noise, etc. But analog circuits are orders of magnitude smaller than digital circuits. So for neural networks of sufficient using analog circuits, may be the only practical hardware implementation. In this work, we desire to use a general purpose output feedback control scheme with the controller learning the approporate control law in real-time. For a feed forward neural network, the scheme depicted in Figure 1 provides the capabilities we desire. ANN Controller Output Control signal

DOI: 10.1109/ISCAS.1999.777586

Extracted Key Phrases

6 Figures and Tables

Cite this paper

@inproceedings{Liu1999FullyPO, title={Fully parallel on-chip learning hardware neural network for real-time control}, author={Jingbo Liu and M. Brooke}, booktitle={ISCAS}, year={1999} }