Allon Guez

Neural network architecture for control

Two important computational features of neural networks are associative storage and retrieval of knowledge, and uniform rate of convergence of network dynamics independent of network dimension. It is indicated how these properties can be used for adaptive control through the use of neural network computation algorithms, and resulting computational advantages are outlined. The neuromorphic control approach is compared to model reference adaptive control on a specific example. It is shown that the utilization of neural networks for adaptive control offers definite speed advantages over traditional approaches for very-large-scale systems.<<ETX>>

On the stability, storage capacity, and design of nonlinear continuous neural networks

The stability, capacity, and design of a nonlinear continuous neural network are analyzed. Sufficient conditions for existence and asymptotic stability of the networks equilibria are reduced to a set of piecewise-linear inequality relations that can be solved by a feedforward binary network, or by methods such as Fourier elimination. The stability and capacity of the network is characterized by the postsynaptic firing rate function. An N-neuron network with sigmoidal firing function is shown to have up to 3/sup N/ equilibrium points. This offers a higher capacity than the (0.1-0.2)N obtained in the binary Hopfield network. It is shown that by a proper selection of the postsynaptic firing rate function, one can significantly extend the capacity storage of the network. >

On the solution to the inverse kinematic problem

A nonalgorithmic method is presented for the solution to the inverse kinematic problem of a robot. The method is robot independent and involves a hybrid approach whereby a neural solution is augmented with an iterative procedure which provides the final solution within some specified tolerance. Essentially, the neural solution is similar to a lookup table in providing a good initial guess to a classical iterative search. It has been found that for the industrial manipulator PUMA 560, the proposed hybrid method achieves about a twofold increase in computational efficiency with better uniformity of the time required to obtain the solution to the robotic manipulator.<<ETX>>

Nonlinear adaptive control with parameter estimation of a CSTR

Abstract The goal of this paper is to describe a linearizing feedback adaptive control structure which guarantees high quality regulation of the output error in the face of unknown parameters. The effectiveness of this control structure is demonstrated on a continuous stirred tank reactor in two instances. The first is when there is full state feedback and the second when only temperature measurements are available. In the latter a nonlinear observer is constructed to infer conversion. In both cases conditions for asymptotic stability are presented and discussed.

Adaptive Performance Control of Computing Systems via Distributed Cooperative Control: Application to Power Management in Computing Clusters

Advanced control and optimization techniques offer a theoretically sound basis to enable self-managing behavior in distributed computing models such as utility computing. To tractably solve the performance management problems of interest, including resource allocation and provisioning in such distributed computing environments, we develop a fully decentralized control framework wherein the optimization problem for the system is first decomposed into sub-problems and each sub-problem is solved separately by individual controllers to achieve the overall performance objectives. Concepts from optimal control theory are used to implement individual controllers. The proposed framework is highly scalable, naturally tolerates controller failures and allows for the dynamic addition/removal of controllers during system operation. As a case study, we apply the control framework to minimize the power consumed by a computing cluster subject to a dynamic workload while satisfying the specified quality-of-service goals. Simulations using real-world workload traces show that the proposed technique has very low control overhead and adapts quickly to both workload variations and controller failures.

Neuromorphic architectures for fast adaptive robot control

An architecture for an adaptive neuromorphic system designed to control a robot is suggested. The proposed architecture utilizes two important features of neural networks: the abundance of local minima in the networks state space and the uniformity of convergence of these minima in the face of growing dimensionality. The proposed approach is expected to yield controllers which are both faster and simpler than controllers which are designed by the methods of model reference adaptive control and self-tuning regulator. The controllers complexity is expected not to grow exponentially with the number of unknown parameters, and to allow adaptation in both continuous and discrete parameter domains. The possible benefits of the architecture are demonstrated on a single-degree-of-freedom manipulator, whose controller is assisted by a neural estimator.<<ETX>>

Neurocontroller design via supervised and unsupervised learning

In this paper we study the role of supervised and unsupervised neural learning schemes in the adaptive control of nonlinear dynamic systems. We suggest and demonstrate that the teachers knowledge in the supervised learning mode includes a-priori plant sturctural knowledge which may be employed in the design of exploratory schedules during learning that results in an unsupervised learning scheme. We further demonstrate that neurocontrollers may realize both linear and nonlinear control laws that are given explicitly in an automated teacher or implicitly through a human operator and that their robustness may be superior to that of a model based controller. Examples of both learning schemes are provided in the adaptive control of robot manipulators and a cart-pole system.

ART based adaptive pole placement for neurocontrollers

Abstract Indirect adaptive control of low order plants that are subjected to parametric variations arising from changes in operating environment requires real time dynamic system identification. In this paper, we propose a control scheme that utilizes a nearest neighbor search type of classifier capable of learning to dynamically identify these variations in plant parameters. The neural network architecture employed is based on the Adaptive Resonance Theory (ART-II) proposed by Carpenter and Grossberg (1987a, 1987b, 1987c, 1987d). An adaptive pole placement controller for a slow time varying linear second order system is implemented based upon this architecture to assess the performance of the network and the overall control scheme with the neural network in the control loop. The control strategy is based upon identification of changes in the time response characteristics of the system to standard test signals which are assessed by the network. A pole placement algorithm is utilized to relocate the poles of the overall closed loop system by altering the gains of the process controller to obtain desired system response. Experimental studies on a simulated system employing a Proportional Derivative controller are encouraging.

Neural networks and fuzzy logic in clinical laboratory computing with application to integrated monitoring

We present an analysis of the computational features of neural networks and fuzzy logic architectures which attempts to explain their recent popularity as well as their drawbacks. Based upon many reports in several fields, we identify the key computational requirements in the clinical laboratory setting, and review several classical tools. In particular we make the observation that all of these needs may be viewed as a search for an appropriate mathematical mapping. We suggest that the neural networks promise as a universal function approximant is the main source of its apparent attractivity. We then describe a customized neural network architecture as a non-linear, adaptive signal processor for integrated monitoring. This architecture is employed in the Adaptive Real-Time Anesthesiologist Associate (ARTAA) system, which has been developed as a joint project at the Department of Anesthesiology, Albert Einstein Medical Center and the Electrical and Computer Engineering Department, Drexel University in Philadelphia, USA. In this application the neural network realizes a non-linear scalar map from the set of physiological signals to a vital function status (VFS) indicator. The system is now under clinical testing.

Online identification and control of linearized aircraft dynamics

An approach for online identification and control that requires weaker excitation than the existing approaches based on least-squares schemes and closed-loop systems is examined. It uses multiple-objective optimization theory to resolve the conflict between identification and controller performance as they compete for the only available resource, the inputs to the aircraft. The approach is applied to a longitudinal model of a representative linearized high-performance aircraft model. Simulation results compare the final controller with a conventional gain-scheduled pitch command augmentation system. It is demonstrated that by allowing some control input to be given to the identification process, the controllers overall performance is improved. >