John J. Helferty

A neural network learning strategy for the control of a one-legged hopping machine

Results are presented on two neural network strategies for the control of dynamic locomotive systems, in particular a one-legged hopping robot. The control task is to make corrections to the motion of the robot that serve to maintain a fixed level of energy (and minimize energy losses), which yields a stable periodic limit cycle in the systems state space. Control of the robot is achieved by the use of artificial neural networks (ANNs) with a continuous learning memory. Through continuous reinforcement for past successes and failures, the control system develops a stable strategy for accomplishing the desired control objectives. The results are presented in the form of computer simulation that demonstrate the ability of two different ANNs to devise proper control signals that will develop a stable hopping strategy, and hence a stable limit cycle in the robots state space, using imprecise knowledge of both the current state and the mathematical model of the robot leg.<<ETX>>

Adaptive Control Of A Legged Robot Using A Multi-Layer Connectionist Network

A Multi-layer Connection Network (MCN) is to control a one-legged mobile robot. The network has no knowledge of the dynamics of the robot, and learns to develop a contol strategy through trial and error. Our results are presented in the form of computer simulations that demonstrate the ability of the (MCN) to devise a set of proper control signals that will develop stable running on a flat terrain.

Experiments in adaptive control using artificial neural networks

A method for using artificial neural networks (ANNs) to control nonlinear, multi-input/multi-output dynamical systems with unknown dynamics is investigated. A neuromorphic controller (NMC) and its application to a nonlinear self-tuning regulator problem is discussed. The NMC performs functions similar to those of adaptive controllers discussed in modern control theory, with the controller taking the form of a nonlinear network and the adaptable parameters being the synaptic interconnection strengths between neurons. The NMC is used to learn a model of the unknown system and to generate the control signals given both the measurements of the current states and the desired values of the current states. The model dynamics is represented by a set of tunable connection weights of the ANN whose weights are adjusted sequentially by a nonlinear recursive-least-squares (NRLS) algorithm which minimizes the error between the desired and current plant states. In effect, the NRLS algorithm trains the ANN to construct mappings of the current state of the plant to the control actions required to maintain the output of the plant at a prespecified value or along a desired trajectory

Neuromorphic control of robotic manipulators

A simple decentralized neuromorphic controller (NMC) for multijoint robotic manipulators with unknown dynamics is presented. The control scheme is computationally very fast and amenable to parallel processing implementation. The NMC is employed to generate the proper control torques to achieve a desired trajectory for the manipulator given both the measurements of the current states and the desired values of the current states. The NMC parameters are adjusted online in real time by the popular backpropagation algorithm which minimizes the error between the desired and current plant states. This method is illustrated by several examples where a two-degree-of-freedom robotic manipulator is controlled to a desired trajectory under a variety of conditions.<<ETX>>

Neuromorphic control as a self-tuning regulator

It is demonstrated that artificial neural networks can be used for the direct adaptive control of both discrete- and continuous-time nonlinear, multi-input/multi-output dynamical systems with unknown dynamics. A neuromorphic controller is presented. and its application to a self-tuning regulator problem is demonstrated. The neuromorphic controller performs functions similar to those of adaptive controllers discussed in modern control theory, with the controller taking the form of a nonlinear network and the adaptable controller parameters being the interconnection weights between neurons. In the discrete-time case the weights are adjusted by a nonlinear recursive least squares (NRLS) algorithm, and in the continuous-time case the weights are adjusted by a backpropagation algorithm.<<ETX>>

Direct adaptive control of flexible space structures using neural networks

A method for using artificial neural networks for the direct adaptive control of large space structures (LSS) with partially unknown dynamics is investigated. A neuromorphic controller is developed, and the authors demonstrate how it is applied to the vibration suppression problem for LSS. A key result is that measurements of all of the system states are not required, but rather only the output measurements and their delayed values. The neuromorphic controller (NMC) is represented by a fixed topology feedforward neural network whose weights are adjusted in real-time by a nonlinear recursive least square algorithm. Several simulation examples are given for the problem of vibration suppression for a subsystem model of the Jet Propulsion Laboratory/AFAL flexible spacecraft simulator.<<ETX>>

Development of a recursive zero annihilator periodic (ZAP) controller with specific applications in flexible space structures

When continuous-time systems are discretized in the digital controller design process, it is often the case that unstable discrete-time zeros (i.e., zeros outside the unit circle in the Z- plane) result regardless of whether or not there are unstable zeros in the original continuous- time plant. Such a system is recognized as being nonminimum phase. Unfortunately, many design techniques in adaptive control are dependent upon pole-zero cancellations and stable plant invertibility and, therefore, cannot be utilized when the plant is nonminimum phase. In this research, a matrix parameter recursive least squares adaptation law is developed for the zero annihilator periodic (ZAP) controller first introduced by Bayard and later extended by Jakubowski. This direct adaptive control scheme allows for the construction of an optimal set of matrix controller gains that place the transmission zeros of the system at the origin, alleviating the nonminimum phase condition, and force the system output to track a desired reference signal. Simulations are presented that demonstrate the performance of the adaptive ZAP controller on a 12-state, 2-input, 2-output partial model of one of the Astrex struts, where the model of the particular strut exhibits nonminimum phase characteristics.

Stabilization of flexible structures using artificial neural networks

A method of using artificial neural networks to stabilize large flexible space structures is presented. The neural controller learns the dynamics of the structure to be controlled and constructs a control signal to stabilize structural vibrations. The network consists of a three layer feedforward network; the input layer receives the displacement and velocity information from sensors located at various points in the structure, and the output layer generates control signals that are applied to the structure through suitable actuators. Sequential updating of the network weights continues, forcing the structure follow a trajectory that eventually leads to complete stabilization. Simulation results on the stabilization of a flexible beam are presented.

Neuromorphic Solution of Nonlinear Two Point Boundary Value Problems

We present a general neuromorphic procedure for solving a class of dynamic programming problems, viz., Two-Point-Boundary-Value-Problems (TPBVPs). The method consists of transformation of the TPBVP to the problem of minimization of an error function over a field of scalars. This error function is then mapped into the energy function of the Tank-Hopfield neural network leading to the synaptic interconnection weights and input bias currents which are adapted to the TPBVP to be solved. The method is illustrated by two examples.

On-line parameter identification and model validation for distributed parameter models with applications to solar thermal systems

The parameter identification problem for a class of distributed parameter plug flow models characterized by first-order vector partial differential equations is investigated. It is assumed that the physical process is represented by a system of partial differential equations of known form but containing unknown parameters. A sequential parameter identification algorithm is developed to estimate the parameters for two distributed-parameter plug flow models, namely the one- and two-temperature-plug-flow models that are used to describe the dynamics of a solar collector. The problem is put into the general framework of a nonlinear identification problem which is solved by a nonlinear recursive least-square (NRLS) algorithm. The NRLS algorithm is used to estimate the model parameters from both stimulated and experimental data taken from Colorado State Universities Solar House III.<<ETX>>