A.V. Sebald

Minimax design of neural net controllers for highly uncertain plants

This paper discusses the use of evolutionary programming (EP) for computer-aided design and testing of neural controllers applied to problems in which the system to be controlled is highly uncertain. Examples include closed-loop control of drug infusion and integrated control of HVAC/lighting/utility systems in large multi-use buildings. The method is described in detail and applied to a modified Cerebellar Model Arithmetic Computer (CMAC) neural network regulator for systems with unknown time delays. The design and testing problem is viewed as a game, in that the controller is chosen with a minimax criterion i.e., minimize the loss associated with its use on the worst possible plant. The technique permits analysis of neural strategies against a set of feasible plants. This yields both the best choice of control parameters and identification of that plant which is most difficult for the best controller to handle.

On the design and performance evaluation of adaptive fuzzy controllers

An automated design technique to optimize membership functions for a fuzzy controller numerically is discussed. The technique is applied to the design of a fuzzy controller for regulating mean aerial blood pressure in a noisy environment (e.g. during surgery or while a patient is being treated in the ICU). The result is extended to construct an adaptive fuzzy controller, and its performance is evaluated through simulation.<<ETX>>

A dynamic empirical model of the human response to sodium nitroprusside during cardiac surgery

A description is given of a partially successful attempt at using real operating room data to generate a dynamic model of the human response to sodium nitroprusside (SNP) during cardiac surgery. The principal goal of the model is to understand the problems associated with adaptive closed loop control of blood pressure during such surgeries. If successful, such a model would be valuable for simulating patients while designing controllers and would also be useful in the construction of estimation algorithms for use in adaptive controllers. The model presented dynamically relates mean arterial pressure (MAP) as a function of SNP infusion and unmeasured disturbances. It uses data (0.5-Hz sampling rate) on MAP and SNP taken over a complete cardiac surgery. Special attention is given to modeling the noise characteristics of the MAP signal.<<ETX>>

An adaptive fuzzy controller for blood pressure regulation

An improved adaptive mechanism suitable for a fuzzy controller is presented. The approach is based on an expert rule base, and the resulting adaptive controller is a knowledge-based object-evaluation fuzzy controller. This control strategy was applied to blood pressure regulation and provided satisfactory performance in simulations. The use of a memory bank for past control history in the adaptive mechanism appears to be an effective way to give the controller benefit from its past experience, providing a rudimentary form of humanlike experience.<<ETX>>

On Making Problems Evolutionarily Friendly - Part 1: Evolving the Most Convenient Representations

The idea of evolutionary friendliness recognizes that problem representations have a significant impact on the performance of evolutionary algorithms. There are two aspects of these representations. Different solution schemes exploit different natural symmetries. Very commonly, problems also possess symmetries that are determined by the coordinate systems used to represent them. Solution symmetries are typically specified by the user and are not allowed to evolve. The problem coordinate system is again typically chosen by the user and not evolved. In this first paper, the most appropriate solution symmetry is evolved. In the second paper, the coordinate system is evolved. In this paper, common detection problems with decision boundaries that possess special symmetries are solved using an evolutionary programming (EP) framework that is capable of exploiting these symmetries to quickly generate solutions. In particular, neural networks possessing appropriate symmetries are evolved by optimizing both their bases and their parameters. Simulation results indicate that the EP procedure is capable of selecting appropriate basis functions for different regions of the input space as well as optimizing the associated set of parameters.

Engineering implications of closed-loop control during cardiac surgery.

Conclusions1Closed loop control in cardiac surgery is a very difficult problem because of changes in patient response, substantial nonrandom external disturbances, random disturbances, and relatively long pure delays between infusion by the pump and initiation of bodily response.2System identification models based on real data are crucial for quantitatively describing the relationships encountered in practice.3A minimax approach is useful for mathematically describing the control design process as an attempt to choose a single controller capable of coping with a set of patients. It also yields computer-aided design tools.4Typical real controllers in complex environments like cardiac surgery contain relatively large numbers of parameters that must be specified. A computer algorithm capable of performing minimax analysis on such problems is available [4].5Design of controllers for complex environments such as cardiac surgery requires the use of computeraided design tools for proper performance. The complexity of the problem and the number of possible combinations of control parameters and architecture may make less exhaustive manual simulation approaches unproductive.6Computer-aided design tools are extremely important because they permit much more exhaustive testing prior to application on animals and humans. They permit evaluation of literally millions of combinations of patients and controllers at computer speeds. Manual design searches cannot be done for these large problems.

Use of neural networks for detection of artifacts in arterial pressure waveforms

Two neural net architectures are applied to the problem of detecting artifacts in pulsatile pressure waveforms emanating from catheters in anesthetized patients awaiting cardiac surgery. A three-layer back-propagation network with 21 nodes in each layer satisfactorily detected all artifacts and falsely characterized none of the patterns. A competitive learning network, although easier to build and train, did not perform nearly so well. In both cases, the networks were trained on one set of data and tested on a different set. Both sets of data were taken from an anesthetized patient about to undergo cardiac surgery.<<ETX>>

Use of simulated annealing in design of very high dimensioned minimax adaptive controllers

This paper presents a practical algorilhm for performing simulated annealing (SA) on minimax adaptive control designs. Techniques for automating SA parameter selection and reducing computer time are presented. ReSultS for several experiments including optimal Selection of membership functions for a fuzzy blood pressure controller are given.

Computer aided design of closed-loop controllers for biomedical applications

A description is given of the use of simulated annealing to optimize closed-loop control designs for biomedical applications. Computer-aided optimization is crucial to the search for useful controllers since many parameters must be chosen to optimize typical controllers. Computerized design tools will permit evaluation of whole classes of algorithms and greatly assist in understanding whether performance problems are due to the basic architecture of an algorithm (e.g. self-tuning PID) or to poor choice of controller parameters. An algorithm is described which is currently running on the UCSD CRAY and on SUN 47260 workstations. The authors have completed design of a robust optimal control and are currently studying a self-tuning regulator and fuzzy controller, all initial examples of the technique based on the problem of arterial blood pressure control with sodium nitroprusside. The real promise of the technique is associated with more complex drug controllers.<<ETX>>

Minimax design of CMAC encoded neural network controllers using evolutionary programming

The authors describe the use of evolutionary programming for computer-aided design and testing of cerebellar model arithmetic computer (CMAC) encoded neural network regulators. The design and testing problem is viewed as a game in that the controller parameters are to be chosen with a minimax criterion, i.e. to minimize the loss associated with their use on the worst possible plant parameters. The technique permits analysis of neural strategies against a set of plants. This gives both the best choice of control parameters and identification of the plant configuration which is most difficult for the best controller to handle.<<ETX>>