Journal of Systems Integration, 5, 187-199 (1995) 9 1995 Kluwer Academic Publishers, Boston. Manufactured in The Netherlands.
Integrated Systems I: Design Principles E A. D. DE MAINE Department of Computer Science and Engineering, 108 Dunstan Hall, Auburn University, AL 36849
K. D. BRADLEY* Department of Computer Science and Engineering, 108 Dunstan Hall, Auburn University, AL 36849 W. H. CARLISLE Department of Computer Science and Engineering, 108 Dunstan Hall, Auburn University, AL 36849 W. B. DRESS
The Oak Ridge National Laboratories, Oak Ridge, TN 37831 Received February 15, 1991; Revised January 19, 1995 Abstract. The following design principles are being used in an ongoing project to realize an integrated family
of rule based systems that can be easily used separately or together in different combinations to solve problems common to many different disciplines. Some essential features of this family are: (1) Individual members can be used in the normal way as user-friendly rule based systems or they can be transparently invoked by other user-friendly rule based systems without interrogating users. (2) The knowledge (or rule) bases of key members do not mimic the perceived mode of human thought; therefore, they can predict events that cannot be predicted by the state-of-the-art alone. (3) The Law of Conservation of Mass/Energy is used to detect and correct computational errors. Keywords: Integrated Systems, Multi-Tier Interfaces, Discipline Independent Rule Based Systems, Autodeduc-
tive Systems, Autolearning Systems, Detection and Correction of Computational Errors
1.
Introduction
The physical and engineering sciences are essentially applications of subsets of the rules for combinatorial mathematics and classification theory. The rules for interpreting results and the various t e r m i n o l o g i e s are characteristic of each discipline. Four p r o b l e m s c o m m o n to m a n y disciplines are: 1.
Manipulation o f graphs, examples of which include c o m p u t e r assisted synthesis o f c h e m i c a l c o m p o u n d s in which the graphs represent c h e m i c a l entities, m e d i c a l diagnoses, and the designs for V L S I and computers.
2.
Evaluation o f graphs where the nodes designate state and edges transformations. Systems o f equations and parsing trees are examples.
3.
Classification. E x a m p l e s include the determination of controlling parameters and the classification of documents.
* Present address: Nichols Research Corporation, Huntsville, AL 35815-1502
188
P.A.D. DE MAINE,ETAL.
Engineering and Science Rule Based Systems Package Automatic Recognition (AUTOREC)
Automatic Learning (AUTOLRN)
Curve-Fi~.ing(CURFIT) KarmarkerAlgorithm(KARMAR) DataCompression(INTEGRAL) DataDescription(JOBLIST) HighSpeedInfoMgmt(SOLID) LiteratureSearch(REFFILER) QueryEnglishDocs.(QED) ProgramValidation(QUEST) ProgramTransportation(TPL)
Manipulating Graphs (SCANCHEM) Evaluating Systemsof Equations (FRANS)
]
Figure 1. Relationships between the component Rule Based Systems for the planned Engineering and Science Ruled Based SystemsPackage. The nine systemsin the central box are the Artificial IntelligenceSystem Support Tools (AISST)each of which can be used on a stand-alonebasis. SOLIDis a high-speedinformationmanagement system. QUEST [3] and SCANCHEM[4] are the work of groups headed by Dr. D. B. Brownand Dr. Z. S. Hippe respectively. REF FILER is a propriety package that can be obtained from SOFT FOCUS. The functions of the individual componentsare briefly discussed in [1], [2].
. Recognition. In the classic pattern recognition problem, one either identifies objects with specific characteristics or determines objects that have unspecified common characteristics.
In order to solve these problems, an Engineering and Science Rule Based Systems Package has evolved consisting of four Special Purpose Rule Based (SPRB) systems and nine Artificial Intelligence System Support Tools (AISST). As the systems are discipline independent they can be used separately or in combinations to solve class problems common to many disciplines [1], [2] (See Figure 1). The four SPRB systems (SCANCHEM, FRANS, A U T O L R N and AUTOREC) solve the above four class problems in the order enumerated. AUTOREC and AUTOLRN are adaptive learning systems whose performance improves with use. S C A N C H E M and FRANS are autodeductive systems that can predict unprecedented events (e.g. events that are not currently in the state-of-the-art). The nine AISST can be used either on a stand-alone basis as user-friendly rule based systems or in combinations with one or more SPRB and other AISST. They perform special tasks like curve-fitting, high-speed information management, data compression, etc. All thirteen systems have both interactive user-friendly interfaces and system interfaces. System Interfaces can be invoked by other rule based systems without interrogating users [5], [6]. At present we have prototypes for ten systems and six of those are in production use. Additional details about these systems are in references [1], [2], [7], [8]. This paper defines the overall design principles that are used to realize major general purpose systems. Details of fully operational systems will be offered in subsequent papers in this series.
INTEGRATED SYSTEMS I: DESIGN PRINCIPLES
189
2. Design Principles The key concepts used to create the Engineering and Science Rule Based Systems Package result from these eight observations: The two mutually exclusive sets of requirements for Man-Machine interfaces, UserFriendly and Machine/System, can be achieved with a multi-tier design in which the user and system interfaces can be independently addressed. . Most problems in the physical and engineering sciences are essentially applications of combinatorial mathematics or classification theory that use subsets of mathematical rules characteristic of the individual disciplines. 3. A modified form of the Law of Conservation of Mass and Energy also applies to mathematical systems of equations. . A simulated telephone network in which the queries themselves describe information path(s) that terminate with locations of the answer(s) becomes a high-speed, logically independent information management system. . Transportation of high-level language code between different machine environments to obtain efficient object code can be achieved with a bifunctional compiler that supports an intermediate hypothetical high-level language that is not compiled. . Scenes, which may have abstract characteristics, can be decomposed into one and/or two dimensional fundamental units, called normalized minimal views, that are the analog of letters in a natural language. 7. The statistical based Salton [9], [10] method of text processing can be significantly improved by using the semantic based "aboutness" concept of Hillman [11], [12] and applications of neural networks to determine candidate terms or classifiers and the content of documents. 8. Parallelism in sequential code can be exploited on high-performance workstations equipped with transputers. Each observation is next discussed in turn.
3.
Design of Interfaces (Observation 1)
Of critical importance is the design of the man-machine interfaces for general rule based systems. They must have an interrogation capability to serve (i) different classes of users from different disciplines and (ii) as service components in integrated packages. These objectives are achieved with a multi-tier interface design whose two parts, the Interactive User-Friendly Interface (IUFI) and the System Interface (SI), can be invoked independently of each other [5], [6] (See Figure 2). The user-friendly interface, which can be altered for different user-groups, passes information to the SI. The SI does not change for new applications and can be invoked by other rule based systems without interrogating users.
190
V.A.D. DE MAINE, ET AL.
1,, INTERACTIVE USER-FRIENDLY INTERFACE 1. User-Friendly and Discipline Dependent. 2. Validity Checks. 3. Passes special instructions to the System Interface. 4. Output compatible with SYSTEM INTERFACE.
0
1. 2. 3. 4.
DATABASE
9
I
SYSTEM INTERFACE Discipline Independent. Validity Checks. Translates to the General Data Structures. Output compatible with MODEL BASE.
MODEL BASE
~
KNOWLEDGE BASE
Figure 2. Multi-tier Interfaces. The rule based system can be invoked by users via the Interactive User-Friendly Interface or transparently by another rule based system via tha System Interface.
Some details of the four different designs that have been implemented in our laboratory are given in PART II of this series [6]. The IUFI should be designed so that it: (1) is interactive, (2) supports a language understood by the user, and (3) is adaptable for new applications. These objectives can be achieved .with a table-driven design with requisite tables designed by experts in the individual disciplines. Other functions that may be included in the IUFI are: (a) checking the validity of inputs with suggested corrective actions, (b) passing special instructions to limit the scope of the model base, (c) ensuring output is fully compatible with the SI, and (d) translating SI output to a format acceptable for user displays. The primary roles of the SI are to convert the input to general data structures used by the model base and to protect the model base against errors received from the IUFI. The SI should be designed to be accessible to the model base (or inference engine), interactive with the IUFI, and able to check and suggest corrective actions if those functions are deferred by the IUFI. This two-tier design for the interface permits the flexibility required by general rule based systems to be used either directly in many different disciplines or transparently by other rule based systems without interrogating users. This design also allows enhancements to user friendliness without affecting the system interface or other related rule based systems.
INTEGRATED SYSTEMS I: DESIGN PRINCIPLES
191
4. Design of Rule Based Systems (Observations 2 and 3) With the Belev definition [13] rule based systems can be classified as Conventional, Autodeductive, Conventional Adaptive Learning, and Adaptive Autolearning. In Conventional expert systems, the separately identifiable rule base is constructed by examining the current state-of-the-art [14]. As such systems ultimately mimic the perceived way humans reason, they cannot predict events outside the current state-of-the-art. In an Antodeduetive system (ADS), the model base is a mathematical model. The rule base is either an integral, indistinguishable part of the mathematical model or it is generated by the system itself from fundamental laws such as the Conservation of Mass/Energy. In such systems uncertainty is entirely due to the accuracy of the data and therefore the concept of "fuzzy logic" should be replaced by the concept of "fuzzy conclusion". Two types of ADS that are used in the science and engineering Rule Based Systems Package are: 1. Numeric Autodeductive Systems, like CURFIT [1], [2] and FRANS [1], [2], [15], are nonsymbolic systems that are "philosophically closed" or "mathematically complete". The rule base is an indistinguishable part of the mathematical model, and there is no database.
2. Alphanumeric Autodeductive Systems, like SCANCHEM [4], are symbolic systems whose rule base is generated by the system itself without reference to either the stateof-the-art or the contents of the data base. The data base contains a list of entities or objects that have been identified in the "real world". Learning systems are used to classify and/or identify entities. Their model bases (or inference engines) consist of one or more clustering algorithms. The algorithms are first used to identify the classification parameters for a known "training set", and then are used to classify unknown sets. Adaptive learning systems augment the training sets with unknown sets that they have previously classified. The classification parameters are continually recomputed, thereby improving their predictive capabilities. Two kinds of adaptive learning systems [1], [2] are: a.
Conventional Adaptive Learning systems, which use manually classified training sets, cannot predict events that cannot be determined from the state-of-the-art.
b.
Adaptive Autolearning systems classify their training sets, are not bound by the stateof-the-art, and are able to predict unprecedented events.
4.1. Designing an Adaptive Autolearning System. The adaptive autolearning system, AUTOLRN, will use a combination of the Borko-Bernick Factorization [ 16] and the Williams Discriminant [17] methods in the following step-wise fashion. (1) Determine candidate descriptors for all documents or entities prior to constructing a weighted frequency matrix.
192
P.A.D. DE MAINE, ET AL.
(2) Calculate eigen values from the frequency matrix. (3) Identify categories by the ranges of real eigen values. (4) Use the Williams Discriminant method to determine the descriptors for the collection. In this step candidate descriptors can be eliminated and categories can be combined. (5) With the output of (4) repeat (2), (3) and (4) until both methods [16], [17] yield the same descriptors and classification categories. 4.2.
Designing a Mathematical Evaluation System.
In its simplest form the Physical Law of Conservation of Mass/Energy states that: (a) equations must be balanced, and (b) parameters that designate mass cannot have negative or complex values [19]. A modified form of this law is that equations must be balanced and parameters can have real or complex values [18]. The realization that the modified form applies to mathematical systems of equations has led to the discovery of: (a) the parameters that control computational accuracy; (b) tests for absolute computational accuracy; (c) the rules for automatically correcting computational errors. These discoveries have led to the development of the fully implemented Error Detection and Corrective Action, EDCA, algorithm for use in solving boundary valued problems. The EDCA is an integral part of FRANS (Figure 1). The user stipulated Computational Accuracy serves to detect mathematical instability and, if possible, to initiate corrective actions until either an acceptable computational accuracy or the limit of computational accuracy is achieved. This method has been used to greatly extend the range of the commonly used Newton-Rhapson and Hamming-Kutta-Rung methods [18], [19], [21]. User stipulated reliabilities of data are used in the Maximum Tolerance Procedure, MTP, to eliminate ambiguities from conventional curve fitting methods [20]. Some details of both the EDCA algorithm and the MTP are given in [18], [19], [20], [21] and especially in [21].
5. Design of General Information Management Systems (Observation 4) The success or failure of systems that require near continuous access to large databases depends on high-speed general information management systems [22], [23]. All attempts to realize applications with large and/or dynamically changing databases appear to be of limited usefulness because they use data management methods whose efficiency is determined by the size and type of database (i.e. they are not logically independent). A High-Speed General Information Management System, HSGIMS, that can be used in any economically viable application must meet the following minimum general specifications. (a) Can support any query language. (b) Data or information independent. (c) Logical (or Question-Type) independent.
INTEGRATED SYSTEMS I: DESIGN PRINCIPLES
193
(d) Very small and bounded search and update times that require at most one access to peripheral storage (i.e. hard disks) for both explicit and non-explicit queries. (e) Has an easily used Security System that can protect the database(s) against viruses and may also be invoked by either users or managers to deny access by unqualified users. (f) Is economical with respect to the use of storage and communications resources. (g) Within the available resources there must be no restriction on the number or size(s) of databases and it must be capable of serving either in a "stand-alone" or distributed environment. (h) It can be transparently used to significantly improve the performance of conventional database systems (e.g. Relational database system). (i) It can be transported for efficient execution in new machine environments. The JOBLIST-SOLID system [24], [25], [26], which fully meets these specifications, processes all queries in less than one access to Direct Access storage plus 0.0001 seconds on a modest machine. It is a simulated communications network in which the queries are descriptions of paths that are to be traced. There is no directory. Because it is essentially an application of Graph Theory (or Abstract Networks) the "internal (or transparent)" query language, JOBLIST, is mathematically complete. Of course that means that the "user (or visible)" query language (which can be a variant of SQL, the proposed SQL3 standard or any other query language) must be translated to the JOBLIST form. Additional flexibility is provided by the nine Override Codes and a Security Mechanism that are integral parts of the Transparent Query Language, JOBLIST. The override codes can be used to formulate queries with virtually any degree of specificity. The central feature of the security mechanism are two different kinds of security lock, Status and Dynamic, that can be inserted in any combination in queries or, for protection against viruses, in any file that is stored on peripheral or secondary storage. Static locks, which are embedded in the simulated communications network and in data files, cannot be easily changed and therefore their location(s) and values can be determined. Dynamic locks are foolproof because they invoke interrogation routine(s) with access to tables that can be easily changed without altering any information in either the simulated communications network or any data file. Of course a key problem is the protection of the small tables that are used by the interrogation routine(s). That can be accomplished with encrypting techniques. The JOBLIST-SOLID system, which is unique in that it fully meets all nine of the general specifications enumerated above, has recently been described in [24], [25], [26].
6.
Design of a Transportable Programming Language, TPL, System (Observation 5)
The TPL system is designed to convert high-level language code from a source environment to an object environment so that it will efficiently execute [27], [28]. A bifunctional compiler installed in the source environment converts the source code into a Hypothetical High-Level Language, HPHLL. The same compiler installed in the object environment converts the
194
P.A.D. DE MAINE, ET AL.
HPHLL to efficient object code for use on the object machine. The advantages of this method are that the transportable programming language system is installed only once in each new environment, there is no restriction on the source or object languages, and programmers do not have to know anything about the HPHLL. This method has been successfully applied to inter-language conversion [27], [28], and it can be used for intra-language conversion (e.g. FORTRAN to ADA).
7. Design of Pattern Recognignion Systems (Observation 6) Virtually all picture processing or pattern recognition systems appear to decompose observations by syntactic, statistical or geometric methods into scenes, and then identify entities in terms of the component scenes [29]-[34]. The problems associated with this approach are: 1. Difficulties in processing entities that have abstract properties, like parts of the signatures of submarines or the properties of classes of functions. 2. Unmanageable databases, even for modest collections. 3. The generally poor performances regarding processing speed and reliability of the results. A different approach, more fully described in [35], will be the basis for the image processing system, AUTOREC. The essential concept is that an observation can be decomposed to sets of non-decomposable patterns of interrelated identifying characteristics called minimal views. Minimal views can be regarded as "letters" that can represent abstractions (e.g. a property of a function or the color of an object) as well as two-dimensional views of an n-dimensional object. An entity is completely defined by an unique combination of minimal views. The immediate consequences are: a. There are provisions for contributions by both abstract and real (or physical) properties of entities. b. Two databases are required. The Minimal View Library is small because a limited number of minimal views describe a large number of different entities. Although the Entity Library, which specifies entities in terms of minimal views, must contain an entry for each different entity, it will be much smaller than conventional databases because minimal views are designated by numbers. C.
The certainty of identification is determined only by the weights assigned to the minimal views in the entity library.
8. Design of Document Classification and Retrieval Systems (Observation 7) This section is included to demonstrate the versatility that is achieved with the two-tier design for interfaces. With respect to both speed of retrieval and the accuracy of answers, all
INTEGRATED SYSTEMS I: DESIGN PRINCIPLES
195
existing major document retrieval systems appear to be inadequate because their database management systems are slow and syntactic methods are used to determine indexes (or descriptors). The problem of determining candidate descriptors or classifiers in natural language text has both semantic and syntactic aspects. The time consuming semantic methods that were pioneered in the DEACON system [36] are not a practical basis for managing large collections of documents. The LEADER system [37] uses philosophically derived techniques to perform high-speed semantic analysis. The most comprehensive of many attempts to assign descriptors by predominantly syntactic methods is found in the SMART system [9], [10], [38]. The basis of the LEADER system is the "aboutness" concept described by Hillman [11], [12]. A high-speed surface-level parser is used to convert English language sentences to simple logical constructs, referential sentences are identified, and the referential terms in such sentences are selected as descriptors. Key problems in this approach are the need for a high-speed data management system to support the libraries and to manage the principal database (a library of condensed text), plus the lack of a natural way to assign weights to descriptors and reduce/expand the descriptor sets assigned to documents or queries. Reduction and expansion of semantic descriptor sets can be achieved by processing the natural language text, before it is parsed, with artificial neural networks in the following ways: 1. An anticipatory system [39] can be used to guess the needs of the user as well as interpret what the user "really means" in case of ambiguous queries. 2. Expectation failure, which occurs when none of the chosen pathways is realized, can be used as the basis for future predictions [40] in a self-correcting fashion. . The context of a word (text) or image fragment (minimal view) is the totality of its proximity relationships with other objects. The meaning of an object resides in the restricted context currently in use. This restricted context is a dynamic subset [41] of the context obtained from the (text or pictorial) database, and changes as the user interacts with the system. As major segments of new knowledge are incorporated into the system, the overall context will need alteration. If symbolic labels are assigned to consistent sets of fragments, a mathematical relationship may be constructed between the two context sets. . The actions that a system produces result from the interaction of the internal states of the system and any external input. In this fashion, behavior imposes a semantics. If this behavior is learnable and modifiable by means of dynamic reclassification of the concept structure [42], the semantics is adaptive in that it changes over time with use of the system. In the SMART system [9], [10], [38], statistical techniques are used to determine candidate words and phrases. Next a novel non-statistical technique is used to convert candidate words to the null stems (i.e. items common to different words) that are the basis of the indexing scheme. Statistical techniques are used to construct concept classes (for expanding
196
a.A.D. DE MAINE, ET AL.
descriptor sets), concept hierarchies (for reducing and expanding descriptor sets), and statistical phrases (to expand descriptor sets). The conventional phrase dictionary mentioned in [38] does not appear to have been implemented. An essential prerequisite for a retrieval system for large scale applications is a high-speed database management system that fully meets the eight general specifications listed at the start of Section 5. The inadequacies (noted above) of the purely statistical (or syntactic) and the purely semantic methods of indexing can be removed by this step-wise procedure in storage operations. a. Use the standard frequency counting methods, with positive and negative dictionaries, to determine candidate terms. b. Determine the referential terms by the Hillman method, augmented by applications of artificial neural networks. c.
Obtain the final list of candidate terms by deleting all terms from the Salton list that are not in the Hillman list and adding those in the Hillman list that are not in the Salton list.
d. Use Salton's suffix list to obtain the null stems for all terms in the final list, obtained in procedure (c). e. Use Salton's statistical methods to determine Concept Classes, Concept Hierarchies, and the two phase dictionaries (statistical and syntactical). In retrieval operations the list of null stems obtained in (e) is augmented with information retrieved from the Concept Dictionary, the Concept Hierarchies, and the two phrase dictionaries. Answers are a ranked list of documents. The "Aboutness" of individual documents will be disclosed by an abstract that displays the key logical constructs in natural language form. The proposed design for a system that will yield a fully integrated package of rule based systems for classifying and retrieving natural language documents is displayed in Figure 3. It consists of a principal system, AUTOTEXT, that will transparently use three Artificial Intelligent System Support Tools, AISST, (JOBLIST, SOLID and QED) and the Special Purpose Rule Based (SPRB) system, AUTOLRN, as service systems. AUTOTEXT is the driver that invokes a service system called QED to determine candidate terms (or the "aboutness") from documents or queries, executes Salton's statistical procedures (in the storage mode) and then, in the retrieval mode, retrieves the relevant documents in their logical construct form. The classification system called AUTOLRN will be invoked by AUTOTEXT itself to perform parts of the statistical analysis and by QED when it classifies documents. QED is the application of Hillman's (augmented with applications of artificial neural networks) and Salton's methods for determining both descriptor terms and the "aboutness" of documents; and the combination Borko-Bernick and Williams method of classifying documents. O'Kane's new implementation of Salton's syntactic method [43] and the high-speed general information management SOLID system [24], [25], [26] will be used through being called by AUTOTEXT and QED.
197
INTEGRATED SYSTEMS I: DESIGN PRINCIPLES
Statistical
AUTOTEXT
AUTOLRN
Retrieve Relevant Documents
Classify Documents
QED Descriptor terms and Classify "Aboumess" of Documents Documents Hillman's ~ w/Neural Networks
Document Retrieval
Borko-Bemiek w/Williams
SOLID (JOBLIST) ~ Retrieval
Figure 3. AUTOTEXT is a proposed rule based system for classifying and retrieving documents. AUTOLRN is a Special Purpose Rule Based (SPRB) system, and QED, SOLID, and JOBLIST are Artificial Intelligence System Support Tools (ASSIST). The Library consists of abstracts of the documents in the form of Hillman's logical constructs.
9. Design of Transputer Systems (Observation 8)
The INMOS Transputers family of microprocessors provides on-chip multiprocessing and communication capabilities. By linking this processor and interface boards to a workstation, one builds a simple multiprocessor system that can communicate both within the system and between workstations using Ethernet connections. Tools such as C, FORTRAN, PASCAL and OCCAM compilers support software development on these systems. Operating system extensions such as Cornell University's TROLLIUS [44] provide the C and FORTRAN functions for internal (in the box) and external (out of the box) communications. Scalability is a benefit of Transputer systems. The TRAM system approach now utilized by SGS Thompson's INMOS division [45] allows new modules to be plugged into existing systems to increase processing capabilities. Prototype systems can thus be easily expanded into production environments. Our goal is to produce systems that can be easily used on any machine configuration with sufficient storage for the application. To reduce the cost and improve accessibility, all developmental work will be done on a network of high-performance workstations, each equipped with transputers. To exploit parallelism, algorithms will be developed to convert transportable sequential code for use on each separate workstation.
198 I0.
P.A.D. DE MAINE, ET AL.
Conclusions
G i v e n are the design principles that are being used to realize a set o f thirteen general rule based systems that can be easily used in laboratories to solve p r o b l e m s c o m m o n to m a n y disciplines. T h e key principles result f r o m observations that: (i) the t w o m u t u a l l y exclusive sets o f requirements for M a n - M a c h i n e Interfaces can be a c h i e v e d with a multi-tier design in w h i c h the user and system interfaces can be independently addressed; (ii) physical and engineering sciences are essentially applications o f combinatorial m a t h e m a t i c s and classification theory that use subsets o f the m a t h e m a t i c a l rules characteristic o f the individual disciplines; and (iii) a modified form o f the fundamental L a w o f C o n s e r v a t i o n o f M a s s / E n e r g y also applies to mathematical systems of equations.
References 1. E A. D. de Maine and M. M. de Maine, "Computer aids for chemists?' Anal Chim. Acta 235, pp. 7-26, 1990. 2. P. A. D. de Maine, "Design principles for discipline independent rule based systems?' Foundations of Computing & Decision Sciences 19, pp. 115-125, 1994. 3. D. B. Brown, K. D. Haga and O. Weyrich, "QUEST---Query utility environment for software testing?' International Test and Evaluation Association 1986 Symposium Proceedings, pp. 38--43, 1986. 4. Z.S. Hippe, "Artificial intelligence in chemistry: Present status and future goals," in Proc. 7th Int. Conf. on Comp. in Chem. Research and Education, Eds. Josef Brandt and Ivar K. Ugi, Dr. Alfred Huthig Verlag GmbH, Heidelberg, West Germany, 1989, pp. 165-178. 5. IF'.A. D. de Maine, B. C. Cartee, M. S. Wojtyna, and M. M. Maine, "A computer tool kit for chemists I. Design considerations for interfaces?' J. Chem. Inf. & Comp. Sci. 30, pp. 155-159, 1990. 6. P.A.D. de Maine and K. G. Price, "Integrated systems IL Multi-tier interfaces for integrating families of systems?' J. of Systems Integration 5, pp. 201-217, 1995. 7. K.D. Bradley and P. A. D. de Maine, "Design and applications of discipline independent rule based systems," in Proceedings 2nd. Expert Systems World Congress, Lisbon, Portugal, 1994. Published on CD-ROM. 8. P.A.D. de Maine and K. D. Bradley, "Development and application of a family of general expert systems," Report No. 18 of the Series: Automatic Systemsfor the PhysicalSciences, Computer Science and Engineering Department, Auburn University, Auburn, AL 36849---40 pages, August, 1990. 9. G. Salton, Automatic Text Processing, the Transformation, Analysis, and Retrieval of Information by Computer. Addison-Wesley: Reading, MA, 1989. 10. G. Salton and M. J. McGill, Introduction to Modern Information Retrieval. McGraw-Hill, Inc.: New York, 1983, pp. 118-145. 11. D. J. Hillman, "Mathematical theories of relevance with respect to the problems of indexing" Report No. 2: "An algorithm for document characterization." Report to the National Science Foundation on Grant No. GN-177, March 12, 1965. 12. D.J. Hillman, "Document retrieval theory, relevance, and the methodology of evaluation." Report No. 1: "Characterization and connectivity." Report to the National Science Foundation on Grant No. GN-451, May 24, 1966. 13. R. H. Be•ev• Kn•wledge Representati•n f•r Decisi•n Supp•rt Systems. N•rth-H•••and Pub•ishing C•mpany: Amsterdam, 1985. 14. R.I. Levine, D. E. Drang, and B. Edelson, "A comprehensive guide to AI and expert systems." McGraw-Hill College Book Division: New York, 1986. 15. P.A.D. de Maine and M. M. de Maine, "Automatic deductive systems I. Chemical reaction models?' Comp. and Chem. 11, pp. 49~65, 1987. 16. H. Borko and M. Bernick, J. Assoc. Comput. Mach. 10, p. 151, 1963. 17. J.H. Williams, Proc. Fall Joint Comput. Conf., p. 161, 1963. 18. P.A.D. de Maine and M. M. de Maine, "Automatic detection and correction of computational errors in programs." J. of Applied Mathematics & Computer Science 4, pp. 101-120, 1994.
INTEGRATED SYSTEMS I: DESIGN PRINCIPLES
199
19. P.A.D. de Maine and M. M. de Maine, "A computer tool kit for chemists III. Error detection and corrective action procedure." Comp. & Chem. 16, pp. 53-60, 1992. 20. P.A.D. de Maine and M. M. de Maine, "A computer tool kit for chemists II. Maximum tolerance procedure" Anal. Chim. Acta 256, pp. 361-368, 1992. 21. P.A.D. de Maine and M. M. de Maine, "Procedures for detecting and correcting computational errors" Report No. 17 of the Series: Automatic Systems for the Physical Sciences, Computer Science and Engineering Department, Auburn University, Auburn, AL 36849, 45 pages, July 1990. 22. D. Bell and J. Grimson, Distributed Database Systems. Addison-Wesley, 1992. 23. J.vandenBurg•``Te•ec•mmunicati•ns-re•atedinf•rmati•nservices•••inPr•ceedings•ftheNAT•Advanced Study Institute on lnforma tion Technology and the Computer Network, Bonas, France, August 21-September 2, 1983. 24. P. A. D. de Maine, K. D. Bradley, and S. M. Jodis, "SOLID: A high-speed logical data independent management system with bounded search times." Online & CD ROM Review 17.1, pp. 1-8, 1993. 25. P.A.D. de Maine, "HSGIMS--A high-speed tool with bound search times for transparently managing and manipulating information." Proceedings of C1KM-93, pp. 216-222, 1993. 26. K.D. Bradley and P. A. D. de Maine, "High-speed general independent management system (HSGIMS)." Proceedings of DATABASE'94, pp. 379-391, 1994. 27. P. A. D. de Maine, S. Leong, and C. G. Davis, "A transportable programming language (TPL) system I. Overview." Int. J. Comp. andlnf Sci. 14, pp. 161-182, 1985. 28. S. Leong, S. M. Jodis, K. Sullivan, O. Jiang, and P. A. D. de Maine, "A transportable programming language (TPL) system II. The bifunctional compiler system." I.E.E.E. Trans. on Software Engineering 16(6), pp. 639646, 1990. 29. R. C. Gonzalez and R. Safabakhsh, "Computer vision techniques for industrial applications and robot control." LE.E.E. Computer, pp. 15-17, 1982. 30. C . S . G . Lee, R. C. Gonzalez, and K. S. Fu, "Tutorial on ROBOTICS," 2nd edition, in I.E.E.E. Computer. Society Press: Washington DC, 1986. 31. R.C. Gonzalez and P. Wintz, Digital Image Processing, 2nd edition. Addison-Wesley Book Express, 1987. 32. K.S. Fu, C. S. G. Lee, and R. C. Gonzalez, Robotics, Control, Sensing, Vision, & Intelligence. McGraw Hill, 1987. 33. T. Y. Young and K. S. Fu, eds., Handbook of Pattern Recognition and Image Processing. Academic Press/Harcourt Brace Jovanovich, 1986. 34. K. Fukunaga, Introduction to Statistical Pattern Recognition. Academic Press Inc., 1990. 35. P.A.D. de Maine, P. A-B. Ng, and M. M. de Maine, "Perception and identification of entities in different environments." Report No. 11 of the Series: Automatic Systems for the Physical Sciences, Computer Science and Engineering Department, Auburn University, Auburn, AL 36849, 35 pages, 1987. 36. J.A. Craig, S. C. Berezner, H. C. Carney, and C. T. Longyear, "Deacon: Direct English access and control." Proc. Fall Joint Computer Conference, pp. 365-380, 1966. 37. D.J. Hillman and A. J. Kasarda, "The LEADER retrieval system." AFIPS Proceedings 34, pp. 4 4 7 4 5 5 , 1969. 38. G. Salton, Ed., The SMART Retrieval System: Experiments in Automatic Document Processing. PrenticeHall: Englewood Cliffs, N J, 1971. 39. R. Rosen, Anticipatory Systems. Pergamon Press: Oxford, 1985. 40. J.H. Andreae, Thinking With the Teachable Machine. Academic Press: New York, 1977. 41. R. Rosen, "Some epistemological issues in physics and biology," p. 322 in Quantum Implications, B. J. Hiley and E D. Peat, Eds. Routledge & Kegan Paul: New York, 1987, p. 322. 42. G.M. Edleman, Neural Darwinism. Basic Books: New York, 1987, pp. 256-69. 43. K.C. O'Kane, "A language for implementing information retrieval software." Online & CD ROM Review 16, pp. 127-137, 1992. 44. K. Boe, S. Dannelly, and J. McCreary, "TROLLIUS--A transputer operating system." Final Report, Computer Science and Engineering Dept., 106 Dunstan Hall, Auburn University, Auburn, AL 36849. 45. TRANSPUTER HANDBOOK--Multiprocessing Made Simple, INMOS Division of SGS, Thompson Microelectronics, October, 1989.