Basic expert system tool

Abstract

A tool for building a knowledge system and running a consultation on a computer is easily mastered by people with little computer experience yet also provides advanced capabilities for the experienced knowledge engineer. The knowledge system includes a knowledge base in an easily understood English-like language expressing facts, rules, and meta-facts for specifying how the rules are to be applied to solve a specific problem. The tool includes interactive knowledge base debugging, question generation, legal response checking, explanation, certainty factors, and the use of variables. The knowledge base language permits recursion and is extensible. Preferably, control during a consultation is goal directed in depth-first fashion as specified by rule order. The tool is easily embodied in assembly language, or in PROLOG to allow user-defined PROLOG functions.

Description

FIELD OF THE INVENTION

The present invention relates generally to artificial intelligence and knowledge engineering, and more particularly to a knowledge engineering tool.

BACKGROUND OF THE INVENTION

Knowledge systems are computer systems that emulate reasoning tasks by using an "inference engine" to interpret encoded knowledge of human experts stored in a "knowledge base." If the domain of the knowledge base, or scope of the problem, is sufficiently narrow and a sufficiently large body of the knowledge is properly coded in the knowledge base, then the knowledge system can achieve performance matching or exceeding the ability of a human expert. In such a case the knowledge system becomes an "expert system."

The most difficult step in building expert systems involves encoding unstructured, often even unarticulated, knowledge into machine readable form. The encoding process is performed by a "knowledge engineer" who must be adept at both the milking of knowledge from a human expert and the encoding of the knowledge into the machine readable expert system language. The ease of the encoding step is dependent on the particular syntax, intelligibility, and capabilities of the expert system language itself, as well as the availability of "knowledge engineering tools" used by the knowledge engineer to test, debug, augment and modify the knowledge base. Due to the lack of knowledge engineering tools based on a transparent expert system language, a person needs a good deal of formal education in computer science as well as specialized training in knowledge engineering to become a skilled knowledge engineer. To build an expert system it is far easier for the knowledge engineer to become a pseudo-expert in the knowledge domain of the human expert than it is for the human expert to learn knowledge engineering and directly encode his or her knowledge into machine readable form.

Another factor limiting the implementation of expert systems is that only a relatively small number of engineers and managers are aware of knowledge systems, so that the number of systems conceived and undertaken is far smaller than it might be. Moreover, after a system is conceived, a small prototype system is desired to demonstrate that a knowledge system would actually be useful in a given context. Currently, some prototyping can be done with well known knowledge engineering tools such as MRS and EMYCIN. Commercial versions of these tools, such as KS300 sold by Teknowledge, Inc. 525 University Avenue, Palo Alto, Calif. 94301, are written in a dialect of LISP and require a rather large computer. The EMYCIN language itself also has a few undesirable limitations such as the inability to handle recursive rules and universally quantified variables.

SUMMARY OF THE INVENTION

The basic object of the invention is to provide a useful knowledge engineering tool for building an expert system and running a consultation on a personal-type microcomputer that is easily mastered by people with little computer experience yet also provides advanced capabilities for the experienced knowledge engineer.

Another object of the invention is to provide a low-cost knowledge engineering tool for the rapid determination of the applicability of knowledge engineering technology to a specific problem domain and rapid prototyping of potential full-scale operational expert systems.

Still another object of the invention is to provide a transparent and easily understood expert system language that supports certainty factors, recursion, symbolic or universally quantified variables, and is extensible.

And yet another object of the invention is to provide a basic expert system tool for a personal-type microcomputer that permits the knowledge engineer to write and add PROLOG subroutines to the knowledge base.

Briefly, in accordance with the invention, a basic tool is provided for building an expert system and running a consultation on a personal-type microcomputer that is easily mastered by people with little computer experience, yet also has advanced capabilities for the experienced knowledge engineer. The basic expert system tool includes a knowledge base in an easily understood English-like language expressing facts, rules, and "meta-facts" which specify how the facts and rules are to be applied to solve a specific problem. The tool includes interactive knowledge base debugging, question generation, legal response checking, explanation, certainty factors, and the use of variables. The knowledge base language permits recursion and is extensible. Preferably, control during a consultation is goal directed in depth-first fashion as specified by rule order unless otherwise specified by the knowledge engineer. The basic expert system tool (hereinafter abbreviated BEST) is easily embodied in assembly language, or in PROLOG to allow user-defined PROLOG functions.

BRIEF DESCRIPTION OF THE DRAWINGS

Other objects and advantages of the invention will become apparent upon reading the following detailed description and upon reference to the drawings in which:

FIG. 1 is a block diagram of the Basic Expert System Tool (BEST) according to the invention;

FIG. 2 is a pictorial diagram of a personal-type microcomputer on which BEST is run;

FIG. 3 is a typical memory map for the microcomputer of FIG. 2 when a PROLOG version of BEST is run on the microcomputer;

FIG. 4 is a hierarchical diagram of the top-level modules or subroutines for BEST;

FIG. 5 is a hierarchical diagram of the lower level modules or subroutines for BEST;

FIG. 6 is a flow chart showing the knowledge base entry syntax for the BEST language;

FIG. 7 is a flow chart showing the BEST language syntax for a fact;

FIG. 8 is a flow chart showing the BEST language syntax for a simple-fact;

FIG. 9 is a flow chart showing the BEST language syntax for a rule;

FIG. 10 is a flow chart of the BEST language syntax for a simple-rule;

FIG. 11 is a flow chart showing the BEST language syntax for a rule premise;

FIG. 12 is a flow chart showing the BEST language syntax for a rule conclusion;

FIG. 13 is a flow chart showing the BEST language syntax for an arbitrary expression;

FIG. 14 is a flow chart of an assembly language subroutine for finding a desired character string in the knowledge base;

FIG. 15 is a flow chart of a subroutine for finding a desired expression in the knowledge base;

FIG. 16 is a subroutine for finding an atom in the knowledge base;

FIG. 17 is a flow chart of a subroutine for finding a knowledge base entry including a desired expression;

FIG. 18 is a subroutine for loading a knowledge base disk file into working memory and stripping multiple blanks and comments during the loading process;

FIG. 19 is a flow chart of a subroutine for interpreting the knowledge base to run a consultation with the user;

FIG. 20 is a flow chart of a subroutine for seeking the value of an expression;

FIG. 21 is a diagramatic representation of cache memory storing the value, certainty factor and reason corresponding to an expression when a value for the expression is determined;

FIG. 22 is a flow chart of a subroutine for trying a knowledge base entry in an attempt to determine the value of a predetermined expression;

FIG. 23 is a subroutine for invoking a rule;

FIG. 24 is a subroutine for testing the premise of a rule;

FIG. 25 is a subroutine for noting the conclusion of a rule or for noting the value of an expression;

FIG. 26 is a flow chart of a subroutine for providing an explanation to the user of why BEST has questioned the user for a value for an expression;

FIG. 27 is a flow chart of the modification to the subroutine in FIG. 14 to permit the binding of arguments to function variables when the knowledge base is being searched for a function;

FIG. 28 is a flow chart showing the modification to the subroutine of FIG. 24 to permit the binding of a variable to the value of an expression in the premise of a rule.

While the invention is susceptable to various modifications and alternative forms, a specific embodiment thereof has been shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that it is not intended to limit the invention to the particular form disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.

DESCRIPTION OF A PREFERRED EMBODIMENT

Turning now to the drawings, there is shown in FIG. 1 a functional block diagram of the preferred embodiment of the Basic Expert System Tool (BEST) residing in a personal-type microcomputer shown in FIG. 2 and interacting with either a knowledge engineer 10 or a consultation user 11. The personal computer generally designated 12 is of conventional construction and has a central unit 13 including a microprocessor and random access memory along with one or more floppy disk drives for receiving computer programs and data to be executed and processed by the microcomputer. To run a program, for example, a floppy disk is inserted in a slot 14 in the front of the main unit 13. In order for the user to interact with the microcomputer 12 when a program is being run, the microcomputer includes a cathode ray tube display 15 for showing the user intermediate results and prompting the user for data to be entered via a keyboard 16. To provide a permanent record or listing of data when a computer program is run, a printer 17 is also provided.

The present invention concerns the use of the microcomputer 12 to design and implement a knowledge system to solve a particular problem. For this purpose, the microcomputer 12 executes a Basic Expert System Tool software module or computer program 18. During execution, the BEST software module 18 resides in the random access memory of the microcomputer 12 and further reserves a portion of random access memory 19 to receive a knowledge base including knowledge about how to solve the particular problem, and a portion of memory 20 called the cache for noting the intermediate results of a consultation.

The Basic Expert System Tool functions in either a building mode or a consultation mode. In the building mode, the knowledge engineer 10 first creates a knowledge base disk file 21 by using a standard text editor program 22. At this time the BEST software module 18 need not be loaded into the random access memory of the microcomputer 12. Once a sufficiently large knowledge base is created in the knowledge base disk file 21, the knowledge engineer 10 tests the knowledge base by loading (23) the knowledge base disk file into the knowledge base random access memory 19. The knowledge engineer 10 then uses BEST 18 to run a test consultation. During the test consultation the knowledge engineer 10 specifies goals for BEST to solve by interpreting (24) facts in the knowledge base 19 and by applying rules of inference in the knowledge base. During the test consultation, conclusions are noted (25) in the cache 20 and BEST looks for prior conclusions (26) in cache.

So that the knowledge engineer 10 may observe the individual steps and internal workings of BEST during the test consultation, debugging commands 27 are entered by the knowledge engineer to expose the internal workings of BEST during the test consultation. BEST generates trace output 28 giving a record or history of precisely how BEST is interpreting the knowledge base 19. Also, the knowledge engineer 10 can stop the test consultation at any given point to change the dynamic state of the consultation being stored in the cache 20. The knowledge engineer can set or reset the cache 20 to any dynamic state. Also, the knowledge engineer 10 can save or transfer (30) the entire contents of the cache 20 to a cache disk file 31 for extensive observation, editing, or to suspend a test consultation at an intermediate point for any length of time. The cache disk file 31 is edited or inspected (32) using a standard text editor. The knowledge engineer 10 can resume the test consultation by loading (33) the cache disk file 31 to the cache random access memory 20. By following this debugging process, the knowledge engineer may discover that minor changes are required in the knowledge base 19. BEST is provided with add/remove commands 34 to permit the knowledge engineer 10 to quickly make minor changes in the knowledge base 19. Once these minor changes are checked out using the debugging commands 27 during a test consultation, the knowledge engineer 10 may enter them into the knowledge base disk file 21 using the standard text editor (22).

In summary, the knowledge engineer 10 creates the knowledge base file and then debugs it with the facilities available in BEST. The knowledge engineer writes facts and rules into the knowledge base disk file 21 with the standard text editor 22. When loaded into BEST, the disk file becomes the knowledge base that BEST accesses whenever it needs information. As BEST pursues a test consultation, the engineer traces the flow of inferences and conclusions BEST makes from the knowledge base and stores in the cache 20. Additional commands allow the knowledge engineer to add or remove entries from the loaded knowledge base, run a test consultation to find the values of goal expressions, set and reset the dynamic state of the consultation in the cache, and continue with the consultation.

During the consultation mode of the Basic Expert System Tool, the consultation user consults (35) with BEST 18 and receives advice (36) on a specific subject. BEST asks the user questions as it makes inferences based on the contents of the loaded knowledge base 19. The user then supplies the required information or in turn asks BEST to reveal why a particular piece of information is needed. The user is free to list (37) knowledge base entries related to a particular aspect of the consultation, as well as to show (38) the current conclusions BEST 18 has noted in the cache 20.

In a specific embodiment of the BEST as shown in FIG. 1, the microcomputer 12 is an IBM (International Business Machines, Inc.) "personal computer" having 128 kilobytes of random access memory. This random access memory is configured as shown in FIG. 3 when the BEST module 18 is loaded and executed. The loading of the disk files 21, 31 as well as the BEST module 18 is performed by the operating system of the microcomputer 12. For the IBM personal computer, the operating system comprises 30 kilobytes of memory (40) containing the standard "MS-DOS" operating system. BEST may use a 14 kilobyte window package (41) for formulating information on the display 15 (FIG. 2) in a particularly attractive fashion. According to the known technique of windowing, multiple listings of data are displayed simultaneously. This is useful for displaying a consultation listing (as shown in Appendix VIII to this specification) on one side of the display 15 and the rules currently under consideration on the other side of the display.

In the specific embodiment shown in FIG. 3, the BEST module 18 is written in a high level programming language called PROLOG. It is easy to implement BEST in PROLOG, since searching, matching of expressions, and instantiating variables are inherent PROLOG operations. In FIG. 3, the inherent PROLOG operations are performed by a PROLOG module 42 residing in 20 kilobytes of random access memory. Also, the PROLOG module 42 uses 8 kilobytes of memory 43 at higher addresses. Since the PROLOG module 42 is used in conjunction with the BEST module 18, BEST is provided with capabilities to allow the user to define his or her own PROLOG functions. Thus, the more experienced knowledge engineer can extend BEST to suit his particular needs by adding user-defined PROLOG subroutines.

During execution of the BEST module 18, random access memory in addition to the cache 20 is needed to store intermediate results. These additional intermediate results are stored in stacks 44 at the upper most 8 kilobytes of memory. In particular, these stacks store pointers for searching the knowledge base 19 for particular expressions, parameters of subroutines that are called recursively, and tables of bound variables and their bindings.

The specific embodiment shown in FIG. 3 uses an 8086 PROLOG-1 dialect manufactured and sold by Expert Systems Limited, 9 Westway, Oxford, OX2 OJB, ENGLAND. The PROLOG language is generally described in the text Programming In Prolog, by W. F. Clocksin & C. S. Mellish, Springer Verlag Inc., New York, N.Y. (1981). A reader familiar with the intricacies of PROLOG will understand that the preferred embodiment of the Basic Expert System Tool further described below advantageously incorporates the inherent feature of PROLOG. For the benefit of the reader who is not familiar with PROLOG, however, the following description will not presume any knowledge of PROLOG on the reader's part, and will show how the Basic Expert System Tool may be programmed in any desired language including assembly language. An assembly language version, for example, would be preferred for small microcomputers that cannot support the PROLOG language.

The language with which the Basic Expert System Tool is written is of some significance. But the language in which the knowledge base is written--the "expert system language"--is the most important factor determining the utility of the Basic Expert System Tool. This is due to the fact that the knowledge engineer 10 does most of his or her work in the expert system language. A knowledge engineer, for example, can use BEST as soon as he or she understands the expert system language regardless of whether he or she understands the language in which BEST is written.

The expert system language of BEST is illustrated by the exemplary "Wine Advisor" knowledge base listed in Appendix VII to the present specification.

For the sake of illustration, an abbreviation of the Wine Advisor knowledge base is reprinted below:

goal=[wine].

color=red.

question(preferred-body)=

`Do you prefer light, medium, or full bodied wine?`.

legalvals(preferred-body)=[light, medium, full].

rule 1: if preferred-body=full then recommended-body=full rule 2: if color=red and recommended-body=full then wine=burgundy. It should be observed that the knowledge base consists of a list of knowledge base entries in the form of EXPRESSION=VALUE. The knowledge base entries are delimited or separated from each other by a period at the end of each knowledge base entry. The first knowledge base entry declares a list of goals for a consultation. For the Wine Advisor the goal is to determine a kind of wine. The second knowledge base entry is a fact stating that the expression "color" has a value "red". The third knowledge base entry declares a question for the expression "preferred-body". The fourth knowledge base entry declares a list of legal values for the expression "preferred-body". The fifth knowledge base entry is a rule declaring that if the expression "preferred-body" has a value "full", then the expression "recommended-body" has a value "full". The sixth and final knowledge base entry is a rule declaring that if the expression "color" has a value "red" and the expression "recommended-body" has a value "full", then the expression "wine" has a value "burgundy".

With the above knowledge base, a consultation proceeds as follows. For each goal in the goal list of the first knowledge base entry, BEST searches for the value of the goal. In this particular case, the expression "wine" is the single goal. BEST searches starting at the top of the knowledge base for the expression "wine". BEST finds the expression "wine" in the first knowledge base entry. This first knowledge base entry, however, does not declare a value for the expression "wine". Thus, searching continues downwardly through the knowledge base for the expression "wine". BEST finds the expression "wine" in the final knowledge base entry, and BEST recognizes that this final knowledge base entry is in the form of a rule which concludes a value for the expression "wine". Thus, BEST applies Rule 2.

To apply Rule 2, BEST evaluates the premise of the rule, which requires the testing of whether the expression "color" has the value "red", and whether the expression "recommended-body" has the value "full". First, BEST searches for a knowledge base entry having the expression "color". BEST finds that the second knowledge base entry declares the expression "color" to have a value "red". Thus, the first conjunctive term of the premise of Rule 2 is satisfied. Next, BEST searches for the expression "recommended-body" in order to determine its value. BEST searches the knowledge base and finds that the expression "recommended-body" is in the conclusion of the fifth knowledge base entry, or Rule 1. This process of recursively evaluating rules in an attempt to test the premise of a rule is called "backchaining." BEST applies Rule 1 and searches for the value of the expression "preferred-body" in the premise of Rule 1. BEST finds that the expression "preferred-body" is included in a question in the third knowledge base entry.

In response to this question declaration, the user is asked the question "Do you prefer light, medium, or full bodied wines?". After the user responds, presumably with the answer "full", BEST checks to see whether there is a legal values declaration for the expression "preferred-body". BEST finds that there is indeed a legal values declaration for this expression, and the list of legal values includes the value "full". Thus, the user's answer is checked to determine whether it is included in the list of legal values. If it was not included, the user would be told to select one of the legal values or to answer with the value "unknown". Since the user has responded with the value "full", the premise of Rule 1 is satisfied thereby assigning the value "full" to the expression "recommended-body". Consequently, the second conjunctive term in the premise of Rule 2 is satisfied, making the premise of Rule 2 true. Hence, Rule 2 finds a value for the goal "wine" equal to "burgundy". BEST concludes the Wine Advisor consultation by stating that "BEST has reached the following conclusions about wine: wine=burgundy (100%) because rule 2." The value 100% denotes that BEST is certain in its conclusion as to the value "burgundy" for the goal "wine".

From this simplified example, it is seen that BEST reasons about symbolic expressions that denote aspects of some situations such as the characteristics of a problem, its cause or its solution. BEST's reasoning is controlled by the list of goals and the order of the facts, rules, and "meta-facts" or declarations such as questions about the problem domain. BEST's basic operation is to find or cumulate evidence for the value of expressions.

The terms "expression" and "value" are entirely general and denote merely a delimited character string in the knowledge base. The only distinguishing feature between an "expression" and "value" is the fact that values become associated with expressions during BEST's interpretation of a knowledge base. Expressions in the BEST language need not have just one value, nor does a value have to be certain. Typically, there will be evidence for several competing values. When reasoning with certain knowledge, however, BEST behaves as an advanced symbolic programming language. BEST can also justify the value it has found for some expression and, during a consultation, explain its current line of reasoning. This transparency, combined with BEST's ability to perform symbolic reasoning with uncertain or incomplete knowledge, makes BEST an ideal tool for building expert systems which mimic a human consultant.

Turning now to FIG. 4 there is shown a hierarchical schematic diagram of the top level software modules or subroutines comprising BEST. Execution starts with a top level module (TL). The top level module calls an initialization module (TL INIT) to initialize the PROLOG module (42 in FIG. 3). Once initialized, the top level module (TL) calls a module (BEST READ) to read in the BEST module 18. Then, the top level module (TL) invokes an execution subroutine (BEST DO) to execute the BEST module 18. When the execution module (BEST DO) is executed, the user is given a top level prompt consisting of a question mark followed by a hyphen, and the system will respond to a set of top level commands that are for convenience listed in Appendix II appended to the specification. These top level commands cause the execution module (BEST DO) to execute one or more of a number of modules including (ADD), (REMOVE), (SET), (RESET), (KBLOAD), and (KBGO). The (ADD) module permits the knowledge engineer 10 to add knowledge base entries to the knowledge base 19 (FIG. 1). The subroutine (REMOVE) enables the knowledge engineer 10 to remove entries from the knowledge base 19. The subroutine (SET) permits the knowledge engineer 10 to set the cache 20 so that any desired expression will have a desired value. Conversly, a subroutine (RESET) is provided to reset the cache 20 to remove desired expressions from the cache 20. A subroutine (KBLOAD) is provided to load the knowledge base from the knowledge base disk file 21 to random access memory 19. In response to the top level command "go" the execution subroutine (BEST DO) calls the subroutine (KBLOAD) and then calls a knowledge base execution subroutine (KBGO).

Turning now to FIG. 5 there is shown a hierarchical schematic diagram of the knowledge base execution subroutine (KBGO) which interprets the knowledge base 19 to conduct a consultation. The subroutine (RESET) is called to clear the entire the cache 20. Then, for each goal expression, a subroutine (SEEK) is called to find the value for the goal expression and a subroutine (SHOW) is called to display the rule or reason which established the value of the goal expression. The subroutine (SEEK) which finds the value for any given expression first calls a subroutine (INCACHE) to determine whether the cache 20 already has been set to the values of the desired expression. If the cache does not include the desired expression, then the knowledge base 19 is searched for knowledge base entries including the desired expression. For each entry including the desired expression, a subroutine (TRY) is called in an attempt to determine the value of the expression from the knowledge base entry.

If the knowledge base entry is in the form of a rule having the expression in its conclusion, a subroutine (INVOKE) is called to apply the rule. A subroutine (TEST) is called to evaluate the premise of the rule. Subroutines (SPECIAL CASES) are called to apply particular rules for logical operations in the premise until the premise is broken down into individual expressions which are evaluated by recursively calling the subroutine (SEEK) which evaluates expressions. If the premise clause of a rule is satisfied, then a subroutine (NOTE) is called to record in the cache the values of the concluded expression.

If the knowledge base entry being tried is in the form of a question for the expression, and the value of the expression is unknown, then a subroutine (ASK FOR) is called to ask the user the question. A subroutine (RESPOND TO) is called to check the user response. If the response of the user is "why", a subroutine (WHY) is called which gives an explanation or calls a subroutine (SHOW) to show the rule which is causing the question to be asked. All of the current rules being applied are shown in response to a user response "show".

If the user response is neither "why" or "show", a subroutine (LEGAL) searches the knowledge base for a legal values declaration for the expression. If such a legal values declaration is found, the legal values are obtained and a subroutine (COMPLETE-RESPONSE) is called in an attempt to match the user's response with one of the legal values. If the response matches the beginning of only a particular one of the legal values, then the user response is completed by replacing it with the matching legal response. Otherwise, if there is no unambiguous partial match, the user is shown the legal values and is requested to respond with a legal value or answer with "unknown" or request further information by responding "why". When a legal value is obtained, the subroutine (NOTE) is called to record the expression and its value. Also, the user may alway append a "certainty factor" to his or her response. The certainty factor indicates the extent to which the user believes the value to be true. A certainty factor greater than 80%, for example, indicates substantial certainty in the value and a certainty factor of less than 20% indicates that the value is unknown. The certainty factor may range from 0 to 100% with a value of 50% indicating that there is some evidence for the expression having the corresponding value. Any certainty factor received from the user is stored along with the value received for the expression. If a certainty factor is not supplied, it is assumed to be 100%. The value, in other words, is then presumed to be known with certainty. The threshold numbers of 20% and 80% are somewhat arbitrary since the certainty factor need not be a precise measure of the probability of the value being true.

In order for BEST to interpret the knowledge base, it is evident that the knowledge must be encoded according to a rather precise format. This syntax is diagramatically illustrated in FIGS. 6-13. The major subdivision of the knowledge base is a KNOWLEDGE BASE ENTRY shown in FIG. 6. Any knowledge base entry may have an optional LABEL identified by a postfix colon. The body of each knowledge base entry is called a FACT. Each knowledge base entry is terminated or delimited by a period. A RULE is a particular kind of fact that is conditioned by a premise.

Turning to FIG. 7 there is shown a flow chart for the entry of a FACT. A fact always includes a SIMPLE-FACT. Optionally, the simple-fact may be post-fixed by the operation "cf" followed by an integer number ranging from 0 to 100 denoting the percentage certainty factor with which the simple-fact is being declared.

Turning to FIG. 8 there are shown the components of a SIMPLE-FACT. The simple-fact is in the form of one EXPRESSION followed by the operator "=" and terminated by another EXPRESSION which is the declared value of the first expression.

Turning now to FIG. 9, there is shown the format for a RULE. A rule always includes a SIMPLE-RULE followed by an optional certainty factor.

Turning to FIG. 10 there is shown the format for a SIMPLE-RULE. The simple-rule is identified by the operator "if" followed by an expression called a "PREMISE" which is in turn followed by an operator "then" and the simple-rule is terminated by a particular kind of expression called a CONCLUSION.

The format for a PREMISE is shown in FIG. 11. The premise always includes at least one SIMPLE-FACT. More than one simple-fact may be joined by the logical infix operator "or" or "and", and each simple fact may be preceded by the logical prefix operator "not". Moreover, complex premises may be built up using left and right parentheses to specify the order of evaluation of the logical operators.

In FIG. 12 it is shown that a CONCLUSION is merely a conjunction FACT's.

Turning to FIG. 13 it is seen that an EXPRESSION is merely any combination of delimited character strings. The smallest delimited character string is either an "ATOM", a "NUMBER", a "VARIABLE", or an "OPERATOR". These character strings are delimited by characters such as a blank, left parentheses, right parentheses, quote, left bracket, right bracket, etc. An atom is a delimited character string starting with a small alphabetic letter. Atoms, for example, represent discrete objects or ideas. A number is a delimited character string starting with a digit. A number, for example, is an integer consisting of a string of digits. A variable is a delimited character string starting with a capital alphabetic letter.

A variable can be instantiated or bound to an arbitrary expression. The VARIABLE-EXPRESSION relationship is distinguished from the EXPRESSION-VALUE relationship due to the fact that a variable may be bound and unbound to different expressions while a value becomes permanently bound or attached to an expression. Variables are used in rules so that the rules have universal applicability to more than one object. For the Wine Advisor, a default rule for determining the value of the atom "recommended-body" could be as follows:

rule 3: if preferred-body=X then recommended-body=X. Thus, presuming that the legal values of "preferred-body" are "light", "medium", and "full", Rule 3 states three rules, one for the variable X being bound to each of the legal values. It should be noted that the character string "preferred-body" is an atom because the hyphen is not interpreted as a delimiter character unless the hyphen is followed by a variable. (The atom preceding the hyphen becomes a function in this case, as described below in conjunction with Appendix X to the specification.) Specifically, the hyphen performs the function of an "atom bond" to improve the legibility of facts and rules.

An OPERATOR is defined as a character string which is given special significance during the evaluation of an expression in terms of the bindings of the variables in the expression and the values of the atoms and numbers in the expression. Certain operators are in the form of atoms and are distinguished from atoms only as certain predefined character strings called reserved words. A list of such reserved words are given in Appendix VI along with other operators. During the seeking of value for an expression, the BEST interpreter searches for operators in the expression. When an operator is found, BEST applies predefined rules constituting the definition of the operator upon the expression in order to simplify or break down the expression into the bindings and values of its component variables, numbers, and atoms. One advantage of implementing BEST in a higher level language such as PROLOG is that the higher level language has already set up a mechanism for decomposing expressions into the component parts and evaluating standard operators in the expression.

The interpretation of the knowledge base is performed by matching the "templates" of FIGS. 6-13 against the character string comprising the knowledge base. Before this matching is attempted, multiple blanks and comments in the knowledge base disk file are deleted. A comment in the knowledge base disk file is identified by the left delimiter "/*" and the right delimiter "*/". The basic operation, then, is finding a predetermined character string in the knowledge base.

For the reader unfamiliar with PROLOG, a typical assembly language program for finding a predetermined character string is shown in FIG. 14. The knowledge base is an expression having lower and higher memory addresses EX1 and EX2, respectively. The character string is defined by low and high addresses C1, C2, respectively, in another part of memory. The result of the subroutine (FINDCHAR) is the character string found in the knowledge base at low and high addresses designated EC1, EC2, respectively. Another return parameter is a logical flag S indicating whether the search was successful (S=true), or unsuccessful (S=false).

The FINDCHAR subroutine generally designated 100 starts with an initialization step 101 wherein a character string pointer IC is set to the low character address C1, and the high return parameter EC2, used as a pointer to the knowledge base, is set equal to the low address of the knowledge base EX1. The logical flag S is initially set equal to false. In step 102 the first character in the knowledge base is compared to the first character in the character string. If the characters do not match, and if there are remaining characters in the knowledge base as tested in step 103, then the knowledge base pointer EC2 is incremented in step 104 so that the next character in the knowledge base is compared to the first character in the character string C1, C2.

Once step 102 indicates a match, the initial match is marked in step 105 by setting the low return address EC1 to the value of the knowledge base pointer EC2. In step 106 the character string pointer IC is compared to the high character string address C2 to determine whether the entire character string has been matched. If so, then in step 107 the logical flag S is set true and execution returns successfully to the calling program. Otherwise, in step 108 the knowledge base pointer is compared to the high knowledge base address to determine whether the entire knowledge base has been scanned. If so, the FINDCHAR subroutine has failed and execution returns to the calling programs. Otherwise in step 109 the character string pointer IC and the knowledge base pointer EC2 are both incremented in order to compare the next character in the character string to the next character in the knowledge base. This comparison is performed in step 110 and if it is successful, execution returns to step 106 to test whether the entire character string has been matched. Otherwise, the match has failed, and in step 111 the knowledge base pointer EC2 must be set back to the next character after the initial match at the knowledge base position having been marked in step 105. Also, the character string pointer IC must be set back to the first character and execution returns to step 102 to test for another initial match.

Turning now to FIG. 15 there is shown a FINDEXP subroutine to find a predetermined expression in the knowledge base. An expression is distinguished from an arbitrary character string in that the expression must be preceeded and followed by either a delimiter character or the beginning or end of the knowledge base. It is presumed that the low address of the knowledge base is KB1 and the high address of the knowledge base is KB2. The low and high address of the predetermined expression are C1 and C2, respectively. The return parameters are EC1, EC2, and the logical flag S as previously defined. The first step in the FINDEXP subroutine generally designated 120 is step 121 wherein the parameter EX1 is set to the low address KB1 of the knowledge base. Then in step 122 the FINDCHAR subroutine of FIG. 14 is called to search for the expression character string in the knowledge base. In step 123 the logical flag S is tested to determine whether a match was found and, if not, the expression is not in the knowledge base, the FINDEXP subroutine has failed, and execution returns to the calling program. Otherwise in step 124 the low address of EC1 the character string in the knowledge base is compared to the low address KB1 of the knowledge base to determine whether the string is at the beginning of the knowledge base. If the string is not at the beginning of the knowledge base, then the character LCH at the left of the character string is obtained in step 125 at the memory address of one less than the low address of the character string. In step 126 this character LCH is compared to each character in the predetermined set of parsing characters and if LCH is not a parsing character, then the initial match was incorrect. If the upper string address EC2 is not less than the upper address of the knowledge base KB2, as tested in step 127, then searching in the knowledge base cannot continue, and logical flag S is set false in step 128, and execution returns to the calling program. Otherwise, in step 129, the initial search pointer EX1 is set to one plus the lower string address in the knowledge base EC1 and execution returns to step 122 to continue searching.

If in step 126 the character LCH was a parsing character, then the upper or right most end of the character string in the knowledge base must be tested to determine whether it is either at the top of the knowledge base or delimited by a parsing character. In step 130 the upper address LEC2 of the string in the knowledge base EC2 is compared to the upper address of the knowledge base. If the string address is not less than the knowledge base address then the expression has been found in the knowledge base and execution returns. Otherwise, in step 131, the character RCH to the right of the string in the knowledge base is taken from memory at an address of one plus the upper address of the character string. In step 132 this character RCH is compared to each character in the set of parsing characters. If RCH is not a parsing character, then searching through the knowledge base must continue and execution jumps to step 129. Otherwise, the expression has been successfully found in the knowledge base and execution returns to the calling program.

Turning now to FIG. 16 there is shown a subroutine for finding an atom in the knowledge base. The knowledge base is searched starting at a low address EX1 and terminating at a high address EX2. The beginning and ending address of the first atom found in the knowledge base is returned by the parameters EA1 and EA2, respectively. The logical variable S is also returned to indicate whether an atom was found. The FINDATOM subroutine generally designated 140 starts with step 141 wherein the return address EA2, which is also used as a pointer, is initially set to the low starting address EX1 and the logical flag S is set false. In step 142 the character in memory at the pointer EA2 is found and in step 143 it is compared to the range of values for small alphabetic letters.

If the character is not a small alphabetic letter, then in step 144 the character is compared to the predetermined set of parsing characters. If the character is not a parsing character, then in step 145 the pointer EA2 is compared to the maximum address EX2 to determine whether there are more characters in the knowledge base to search. If there are not any more characters in the knowledge base to search, the search for an atom has failed and execution returns to the calling program. Otherwise, in step 146 the pointer EA2 is incremented to continue searching for an initial parsing character.

Once a parsing character is found in step 144, then in step 147 the pointer EA2 is compared to the maximum address EX2 to determine whether there are more characters to search. If not, the search for an atom has failed and execution returns to the calling program. Otherwise, the pointer EA2 is incremented in step 148 and execution jumps back to step 142 in an attempt to find a small alphabetic letter.

If in step 143 the character is found to be a small alphabetic letter, then the start of the atom is marked in step 149 by setting the low return parameter address EA1 to the current value of the pointer EA2. In step 150 the pointer EA2 is compared to the high string address EX2 to determine whether the upper end of the character string has been reached. If not, the pointer EA2 is incremented in step 151 and the character in memory specified by the pointer EA2 is obtained in step 152. In step 153, this character is compared to a parsing character. If it is a parsing character, then the upper or right most character of the atom has been determined. Before returning, however, the pointer EA2 must be restored to the last character in the atom by decrementing the pointer in step 154. If, however, the pointer EA2 is at the upper end of the string, as determined in step 150, then the pointer EA2 need not be adjusted. Finally, in step 155, the logical flag S is set true and execution returns to the calling subroutine.

Similar subroutines for finding a variable and finding a number are virtually identical to the FINDATOM subroutine 140 in FIG. 16. For a subroutine to find a variable, step 143 tests whether the character is a capital alphabetic letter. To find a number step 143 checks whether the character is a decimal digit.

Turning now to FIG. 17, there is shown a knowledge base entry subroutine KBENTRY generally designated 160, which searches the knowledge base, starting at address KB1 and ending at KB2, for an expression in memory at a low address EX1 and a high address EX2. The subroutine KBENTRY returns the low and high addresses of the knowledge base entry EN1, EN2, respectively which includes the expression.

BEST efficiently searches the knowledge base for expressions since a search pointer table is used to store the addresses where searching left off for various expressions. In other words, during the backchaining of rules, a search for an expression that is found in the conclusion of a rule will be interrupted in the middle of the knowledge base in order to search other entries in the knowledge base for expressions in the premise of the rule. Thus, if the premise of the rule fails, searching for the expression in the conclusion can continue in the middle of the knowledge base rather than, for example, starting back at the beginning of the knowledge base.

In order to determine where the search should begin or continue in the first step 161 the pointer table is searched for an expression that is the same as the expression having the low address EX1 and high address EX2. If the search fails, then searching for the expression in the knowledge base has not previously been interrupted in the middle of a search. In other words, in step 162 the logical flag S is compared to the value false. If S is false, then the search failed and the expression, consisting of a string of characters, is added to the pointer table in step 163. Also, in step 164 the pointer in the table associated with the expression is set equal to the low address of the knowledge base KB1. If, however, the search succeeded as determined in step 162, then in step 165 the preexisting pointer for the expression is obtained from the table.

Now that a starting position has been obtained, in step 165' the subroutine FINDEXP is called (see FIG. 15) with the pointer POINT specifying the low address of the knowledge base for searching. In step 166 the logical flag S is compared to false to determine whether the search failed. If the search failed, then execution returns to the calling program. If, however, the expression was found in the knowledge base, then the delimiters of the knowledge base entry including the expression must be found in order to determine the beginning and ending addresses EN1, EN2, respectively, of the knowledge base entry. In step 167 the return parameters EN1 and EN2 are initially set to the beginning and ending addresses EC1 and EC2, respectively, of the expression in the knowledge base. In step 168, the return parameter EN1 is compared to the low address KB1 of the knowledge base. If the expression starts at the beginning of the knowledge base, then there is no need to search for a left delimiter since the first entry in the knowledge base is being considered. Otherwise, in step 168 the parameter EN1 is decremented and used as a pointer to obtain a character from memory in step 170 which is compared in step 171 to the knowledge base entry delimiter which is a period. If the character is not a period, then scanning to the left continues with step 168. Otherwise, the value for the return parameter EN1 has been properly determined and searching starts for the delimiter on the right of the expression. Thus, in step 172 the return parameter EN2 is compared to the upper most address of the knowledge base KB2. If the return parameter EN2 does not point to the top of the knowledge base, it it incremented in step 173 and used in step 174 as a pointer to obtain a character which is compared in step 175 to the delimiter for the knowledge base entry which is a period. Scanning continues until the delimiter is found or until the top of the knowledge base is reached at which time the proper values have been obtained for EN1 and EN2. Before returning, however, the pointer must be updated in the pointer table. If the return parameter EN2 points to the top of the knowledge base, as detected in step 176, then the expression is removed in step 177 from the table since the entire knowledge base has been searched. Otherwise, in step 178, the pointer for the expression in the table is updated to a value of EN2. Once the table has been updated or the expression has been removed from the table, execution returns to the calling program.

The searching by the KBENTRY subroutine as shown in FIG. 17 presumes that comments have been stripped from the knowledge base. Also, to speed up searching it is desirable that multiple blanks be stripped from the knowledge base as well. This stripping is performed by the KBLOAD subroutine shown in FIG. 18 and generally designated 180. If the first step 181, a buffer is set up for reading from the knowledge base disk file. The buffer, in other words, transfers a track of the knowledge base disk file to random access memory from which the characters may be obtained one character at a time. In step 182 a pointer PNT is set equal to the desired low address KB1 of the knowledge base in memory, and a memory location OLDC is set equal to the ASCII or character code of a blank. The subroutine KBLOAD receives a parameter KMAX denoting the maximum memory address allocated to the knowledge base buffer. In step 183 the pointer PNT is compared to this maximum address KMAX to determine whether there is sufficient random access memory to hold the entire knowledge base received from the knowledge base disk file. If the pointer PNT exceeds this maximum address KMAX, then in step 184 an error message is displayed to the knowledge engineer to indicate that the knowledge base disk file is too big. Then in step 185 the buffer is released and the logical flag S is set false prior to returning to the calling program.

If in step 183 it was determined that the pointer did not exceed the maximum address, then in step 186 the next byte from the buffer is read into a memory location CHAR. If the buffer is found to be empty in step 187, then the entire knowledge base disk file has been loaded into memory. In this case, the maximum address of the knowledge base is obtained in step 188 from the current value of the pointer PNTR. In step 189 the buffer is released and the logical flag S is set true. Execution then returns to the calling program.

If, however, in step 187 the buffer was not empty, then the previous character OLDC and the current character CHAR are both compared in step 190 to the character code for a blank to detect the presence of a multiple blank. If a multiple blank occurs, execution returns to step 183 in order to eliminate the current blank character CHAR. Otherwise, in step 191 the old character OLDC is compared to the character code for a slash (/) and the current character CHAR is compared to the character code for an asterisk (*), in order to detect the beginning of a comment. If a comment is not present, then in step 192 the current character CHAR is loaded into memory at the address specified by the pointer PNTR. In step 193 the pointer is incremented and the new character becomes the old character. Execution then returns to step 183 to read more characters from the knowledge base disk file sequentially into the knowledge base memory.

If a comment is detected in step 191, then the comment is read without transferring data to the knowledge base memory, until the end of the comment is detected. In step 194 the old character is replaced by the new character. In step 195 the next character is read from the buffer. If the buffer is empty as tested in step 196, execution is transferred to step 188 since the knowledge base disk file has been completely transfered to the knowledge base memory with a final comment being deleted. Otherwise, in step 197, the presence of the end of the comment is sensed by comparing the old character to an asterisk and the new character to a slash. If the end of the comment is not reached, execution returns to step 194. Otherwise, the value of the pointer PNTR must be adjusted and the old character OLDC must be reset to the values that they would have had but for the comment. In step 198, the value of the pointer PNTR is compared to the low address KB1 of the knowledge base memory. If the pointer is greater than the low knowledge base address, which is the usual case, the pointer is decremented in step 199. Also, if the pointer is greater than the low address of the knowledge base as tested in step 200, the old character is reset to the character in memory at one less than the pointer as computed in step 201. Otherwise, the old character is reset to a blank in step 202. The decrementing of the pointer in step 199 and the reseting of the old character in step 200 eliminates the initial slash in the left delimiter of the comment. Once the comment has been deleted and the pointer and old character reset to their proper values, execution returns to step 183.

Now that the base level subroutines have been described for the Basic Expert System Tool, the top and intermediate subroutines for conducting a consultation should be easily understood from the following description.

The subroutine KBGO generally designated 210 starts in step 211 by completely reseting or clearing the cache memory 20 (FIG. 1). Then, in step 212, the knowledge base is searched for initial data declarations in the form of "initialdata=L", where L is a list of expressions. For each initial data expression in the list L the subroutine SEEK is called to find the value of the expression. Then, in the final step 213, the knowledge base is searched for a declaration of goal expressions in the form of "goal=L" and for each goal expression in the list L, the subroutine SEEK is called to find the value of that expression, and the subroutine SHOW is called to display the value.

Turning now to FIG. 20 there is shown a flow chart of the subroutine SEEK generally designated 220 which finds the value or values for a given expression, and stores the values in the cache memory. In the first step 221 the expression is compared or matched against a predetermined list of built-in and user-defined propositions. The built-in propositions are defined in Appendices IV and V and include arithmetic propositions, propositions about attributes, and knowledge access propositions. The propositions about attributes include, for example, whether a value for an expression is known, is being sought, or is definite. Knowledge access propositions include, for example, functions to return the value or values of an expression stored in the cache. As a specific example, the value of the proposition "mostlikely(X)" is the value of X with the highest certainty factor. This proposition is used to narrow down the scope of a search. The following selection rule, for example, could appear at the bottom of a "Wine Advisor" knowledge base having a single goal=[best-wine]:

rule:

if mostlikely(wine)=X then best-wine=X. If the expression is a built-in or user-defined proposition, then in step 222 the values of any subexpressions or arguments for the built-in or user-defined proposition are sought and the proposition is executed to obtain the value for the given expression. It should be noted that steps 221 and 222 are rather conventional steps for an interpreter. An interpreter for the BASIC language for example, has such steps for evaluating arithmetic propositions.

If the given expression is not a built-in or user-defined proposition, then in step 223 the cache memory is searched for the expression. If expression has already been noted in the cache, then the subroutine SEEK is finished, and execution returns to the calling program.

If the expression is neither a predefined proposition nor already stored in the cache, then the knowledge base must be searched for an entry including the expression. For this purpose, in step 224 the subroutine KBENTRY (FIG. 17) is called to get the next such entry from the knowledge base.

Upon returning from the subroutine KBENTRY, the the logical flag S is tested in step 225 (FIG. 20) to detect the end of the knowledge base, but before returning the cache is checked in step 226 to determine if a value for the expression was found. If at this time the expression is not found in the cache, a certain value of "unknown" is assigned and stored in the cache for the expression so as to prevent future searches through the knowledge base for the expression.

If the end of the knowledge base was not reached, then in step 227, if the knowledge base entry declares the expression to the multivalued, this fact is noted in step 228 so that the knowledge base will be searched for multiple values. In step 229, the knowledge base entry is matched against templates for a rule, a fact defining the expression, or a question for a value for the expression. If the knowledge base entry is in one of these three forms, then in step 230, the subroutine TRY is called in an attempt to determine the value of the expression from the knowledge base entry. In step 231 the seeking of a value for the given expression terminates if the subroutine TRY finds a value with certainty and multiple values are not being sought for the expression.

It should be recalled that the cache is used to return the values of the expression which are found by the subroutine SEEK. The cache memory 20 is organized as shown in FIG. 21 in terms of a table with four entries for each expression-value pair. These four entries are character strings including an expression, a value, a certainty factor, and a reason. If a value for an expression is determined by a question, the reason is "yousaidso". If the value for the expression is concluded by a rule, the reason is the label of the rule. If the value of the expression is determined by a fact, the reason is the label of the fact. If a value was concluded by more than one rule or fact, the reason for the value is the conjunction of the labels for the rules and facts.

Turning now to FIG. 22 there is shown a flow chart of the subroutine TRY generally designated 250 which attempts to use a knowledge base entry to determine a value for the expression if the expression is single valued, or multiple values for the expression if the expression is multiple valued. A parameter SMV initially set to "single" by SEEK indicates whether the expression was determined to be multiple valued in step 228 of FIG. 20. In the first step 251, the label of the knowledge base entry and the certainty factor are stripped off the beginning and end of the knowledge base entry, respectively. Then, in step 252, the knowledge base entry is compared to a template in the form of "if P then C" to determine whether the knowledge base entry is in the form of a rule and to find the premise and conclusion of the rule. The conclusion is also searched for the given expression. If the knowledge base entry is in the form of a rule and the expression is in the conclusion of the rule, then in step 253 the subroutine INVOKE is called to apply the rule. If, however, the knowledge base entry is not in the form of a rule, then in step 254 the knowledge base entry is compared to a template in the form of "question(E)=Q". If the knowledge base entry is in the form of a question for the expression, then the user is asked the question unless the expression is single valued as tested in step 255 and the expression is in the cache and has a certainty factor greater or equal to 20% as tested in step 256. The user is asked for a value for the expression by calling a subroutine ASKFOR in step 257. If the knowledge base entry is neither in the form of a rule nor a question, then in step 258 the knowledge base entry is compared to a template in the form of "E=V". If the knowledge base entry is in the form of a fact declaring the given expression to have a particular value, then the expression is determined to have the particular value. In step 259 this particular value is recorded in the cache by calling the subroutine NOTE.

Shown in FIG. 23 is a flow chart of the subroutine INVOKE generally designated 260 for applying a given rule. In the first step 261 a subroutine TEST is called to test whether the premise of the rule is satisfied. If the premise is satisfied as determined in step 262, then in step 263 the conclusion is noted in the cache by calling the subroutine NOTE.

Turning now to FIG. 24 there is shown a flow chart for the subroutine TEST generally designated 270. This subroutine matches the premise against templates to strip down the premise into its component parts according to the syntax shown in FIG. 11. The operation of the subroutine TEST is analogous to the execution of propositions in step 222 of FIG. 20. The logical operators "and", "not" and "or", however, are defined in a slightly peculiar fashion in order to determine a certainty factor indicating the degree of success of the premise. Also, the logical operation "or" is given precedence over the logical operation "and". The subroutine TEST breaks down the premise by recursively matching the premise against templates for the logical operations. In the first step 271, the premise is compared to the template, "P1 and P2" in order to remove and process all of the "and" logical operators in the premise. For each of the logical "and" operators, a special case sequence 272 is executed. The subroutine TEST is recursively called to test the first term of the conjunction. If the first term of the conjunction fails, then the entire proposition fails and execution returns. Otherwise, the subroutine TEST is recursively called for the second term in the conjunction. Assuming that there are no variables in the first term in the conjunction, if the second term of the conjunction fails so does the premise so that execution returns. (For the case of variables, the premise may succeed in a number of ways, and each possibility is checked, as described below in conjunction with FIG. 28). If both of the terms of the conjunction succeed, then the certainty factor for the premise is calculated as the minimum of the certainty factor of the rule and the certainty factors of the individual terms in the conjunction. Execution then returns with this combined certainty factor.

In step 273 the premise is matched against a template in the form of "not P1". If a "not" logical operator is found, then the special case sequence in step 274 is executed. The subroutine TEST is recursively called for the term P1. Unless the term P1 is determined with certainty, then the premise fails. Otherwise, the premise fails if the term P1 succeeds, and succeeds if the term P1 fails.

In step 275 the premise is matched against a template "P1 OR P2". If the premise is in the form of a logical disjunction, then the special test sequence in step 276 is executed. The subroutine TEST is recursively called for the first term P1. If the first term succeeds, then the entire premise succeeds. Otherwise, the subroutine TEST is called for the second term P2, and the premise succeeds or fails depending upon whether the second term P2 succeeds or fails.

Once the logical operators are stripped from the premise, the premise is in the general form of "expression=value" and is checked in step 277. Thus, the premise is in the general form which may be processed by the subroutine SEEK. In the special case procedure of step 278, the subroutine SEEK is called to find the values and certainty factors for the expression. Upon returning from the subroutine SEEK, the values and certainty factors for the expression are successively obtained from the cache. Unless one of the values from the cache is equal to the value stated in the premise and the corresponding certainty factor is greater than 20%, the premise fails.

If the premise cannot be recognized then in step 279 the logical flag S is set false to indicate that the premise fails, and execution returns.

Turning now to FIG. 25, there is shown a flow chart for the subroutine NOTE generally designated 280 which updates the cache when the premise of a rule succeeds or when a value for an expression is determined. If a premise of a rule succeeds, then in step 281, the expressions and values are obtained by matching the conclusion to the form of a logical conjunction of facts. In step 282, for each expression-value pair and certainty factor, and the cache is first searched for the same expression and value. If the same expression and value are not in the cache, then the new value for the expression and the certainty factor and label for either the rule or fact are stored in a new entry in the the cache. Otherwise, the new value reinforces a previous determination of the same value. The reinforcement is noted by an increase in the certainty factor. The certainty factor is combined assuming that the certainty factors represent probabilities of independent events. Thus, the combined certainty factor CF2 is calculated according to:

CF2=CF1+CF-CF1*CF/100 It should be noted that the factor of 100 is included because certainty factors are stored as integers in terms of a percentage of certainty. Note only is the certainty factor reinforced when the same value is concluded repeatedly, but all of the contributing reasons are stored in the cache. If the reasons are the same, however, the original reason will suffice and the certainty factor is merely updated. Otherwise, the certainty is updated and the new label is joined as a conjunction with the previous reason.

Turning now to FIG. 26 there is shown a flow chart for the subroutine WHY generally designated 290 which explains to the consultation user why he is being asked for a value for a given expression. In the first step 291, the rules currently being invoked are obtained from the rules in the parameter stack of the INVOKE subroutine. It should be recalled that the INVOKE subroutine is recursively called to backchain rules, and hence there may be a number of rules currently being invoked. If this stack is not empty as tested in step 292, the user is told in step 293 that "your answer to this question will help determine if the following rule is applicable in this consultation". The first rule in the stack is listed. If, however, there are more rules in the INVOKE subroutine stack, as determined in step 294, then other rules are also being considered. Therefore, in step 295, the user is given a list of the labels of these rules after being told that "the following rules are also under consideration".

If the INVOKE subroutine stack is empty, then in step 296 the knowledge base is searched to determine whether the expression is in an initial data declaration. If so, the user is told in step 297 that the expression "is always asked near the beginning of a consultation as it is an initial data expression". If, however, the expression is not in the initial data expression, the knowledge base is searched in step 298 to determine whether the expression is declared as a consultation goal. If so, then in step 299 the user is told that the expression "is one of the goals of the consultation".

If it cannot be determined why the user is being ask for a value for a given expression, then in step 300 the user is merely told that "BEST was asked to find a value" for the particular expression.

With the internal details of BEST in mind, the reader should be able to understand how BEST interprets the "Wine Advisor" knowledge base listed in Appendix VII. The knowledge base includes knowledge about selecting a wine to be served with a particular meal. In general, the wine is selected according to the characteristics of the meal, the preference of the person drinking the wine, and the relative qualities of the wines themselves. Typical consultations are shown in Appendices VIII and IX.

BEST starts a "Wine Advisor" consultation by invoking the subroutine KBGO (FIG. 19) in response to the user entering a top-level command "go" after the high level prompt "?-". KGBO resets (clears) the cache and invokes the subroutine KBENTRY(goal=, goal=GLIST) to find the goal statement "goal=[wine]" at the top of the knowledge base. KBGO then calls SEEK (FIG. 20) to seek the first (and only) goal "wine" in the goal list GLIST. SEEK finds the cache empty and checks whether "wine" is multivalued by searching the knowledge base for the expression "multivalued (wine)" by invoking. KBENTRY("multivalued(wine)", KBEN) and upon returning discovering that "wine" is multivalued. Next SEEK invokes KBENTRY(wine, KBEN) to find the first knowledge base entry KBEN involving the goal expression "wine", which is Rule 46. SEEK tries this first knowledge base entry to find a value for wine by calling TRY("wine", "multivalued", "rule46" P, C). Since "wine" is multivalued, KBENTRY and TRY are iteratively invoked until the knowledge base is found to be empty of remaining knowledge base entries including the expression "wine", and each of these entries are checked for other values of "wine".

In the first invocation of the TRY subroutine (FIG. 22), the knowledge base entry "rule46: etc." is matched against "L: if P then C" identifying the knowledge base entry as a rule and obtaining P="recommended-color=red and . . . recommended-sweetness=sweet" and C="wine=gamay". TRY then calls the subroutine INVOKE("wine, " "multivalued," "rule46", P, C) to apply Rule 46.

The INVOKE subroutine (FIG. 23) breaks down the premise P into individual logical propositions by sequentially matching the premise P against the templates "PR1 or PR2" and "PR1 and PR2", obtaining PR1="recommended-color=red." The individual logical clauses are matched against the template "E=V, CF" to obtain the expression E, value V, and CF for each clause. For the first clause, it is found that PR1="recommended color," V1="red", and CF is `null`. The individual logical clause are tested by calling the subroutine TEST (E, V, CF) for each clause until the premise P of the rule fails by virtue of an AND propositions failing or all of the OR propositions failing. TEST is first called with the argument list TEST ("recommended-color," "red," `null`). SEEK is recursively called for the proposition "recommended-color" which is concluded by Rule 26. Rule 26 includes in its premise the expression "best-color" which is itself sought. In seeking "best-color" Rule 20 is invoked, which includes the expression "main-component" in its premise. The expression "main-component" is an argument for the question knowledge base entry asking the user "Is the main component of the meal meat, fish, or poultry?" Hence, the user is asked for one of the legal values "meat", "fish" or "poultry".

In the consultation of Appendix VIII, the user responds "poultry." This causes the premise of Rule 20 to fail, so that another rule concluding "best-color" is sought. The next such rule is Rule 21, which seeks a value for the expression "has-turkey." Consequently, question(has-turkey) is invoked asking the user "Does the meal have turkey in it?" The user responds "yes", causing Rule 21 to fail. The next rule concluding "best-color" is Rule 23, which requires a value for the expression "has-sauce." During seeking of a value for "has-sauce" question(has-sauce) is invoked asking the user "Does the meal have a sauce on it?" The user responds "yes", concluding a legal value for "has-sauce". Next Rule 23 seeks a value for "sauce" which invokes question-(sauce) asking the user "Is the sauce for the meal spicy, sweet, cream or tomato?" The user responds "sweet." This causes Rule 23 to fail. The next rule concluding "best-color", Rule 24, is invoked and succeeds using values already stored in the cache. The conclusion of Rule 24 sets "best-color" to white cf 50. Execution returns to Rule 46, so that of Rule 24 has the effect of defining "best-color" as multivalued. Consequently, the rest of the rules concluding "best-color" are invoked. Rule 25 is the next rule concluding "best-color" and it fails since "main-component" is known to be poultry. Thus, execution returns to the invocation of Rule 26, which concludes "recommended-color" to be "red cf 80", 80 being the certainty factor of the premise of Rule 26.

Since the value of "recommended-color" is not concluded with certainty, the next rule concluding "recommended-color" is tried. Rule 27 concludes "recommended-color" is "white cf 50". Next, Rule 28 seeks a value for "preferred-color" invoking question(preferred-color) which asks the user "Do you generally prefer red or white wines?" The user responds "white." Hence, Rule 28 fails. The next rule concluding "recommended-color," Rule 29, however, succeeds. Rule 29 concludes "recommended-color" is white cf 20. Since the cache at this point already includes the value of white cf 50, Rule 29 reinforces this previous value. The certainty factors are combined by the subroutine NOTE (FIG. 25) according to the formula:

CF.rarw.CF+CF.sub.new -(CF)(CF.sub.new)/100 thereby increasing the certainty factor for "recommended-color" is "white" to 60. At this point the cache contains the entries shown in FIG. 21.

Execution returns to the invoking of Rule 46, causing the seeking of values for "recommended-body" and "recommended-sweetness." Seeking of values for these expressions causes asking the user the questions "Is the flavor of the meal delicate, average, or strong?"; "Do you prefer light, medium, or full bodied wines?"; and "Do you generally prefer dry, medium or sweet wines?" In response to values obtained from the user, Rule 46 concludes "wine" is gamay cf 79%. "Wine," however, is multivalued as determined by the multivalued(wine) declaration at the top of the knowledge base. Consequently, all of the wine-determining Rules 46-57 are invoked. Since "wine" is a goal of the consultation, the "Wine Advisor" of Appendix VIII lists all of the values for "wine" in order of decreasing certainty factor. This concludes the "Wine Advisor" consultation in Appendix VIII.

The explanation facilities of BEST are illustrated in the "Wine Advisor" consultation of Appendix IX. In response to the user responding "why" to the question prompt ">>", the WHY subroutine (FIG. 26) lists the rule last invoked and the labels of the other rules under consideration.

Now that the reader has a basic understanding of BEST the instantiation and deinstantiation of variables can be explained. The ability of BEST to handle variables in recursive rules enables BEST to handle an unlimited range of problems. Technically speaking, the BEST expert system language is a universal language since any computer program can be translated into the BEST language. This should be evident from the ability of BEST to handle arbitrary expressions, variables, and logical and arithmetic operations. Moreover, BEST is extensible since the knowledge engineer can extend BEST's capabilities, for example, by defining functions. As an expert system tool, however, BEST is designed to perform as an efficient backchaining goal-directed inference engine. Due to this objective, there is one significant limitation on the use of variables. Simply stated, this limitation is that an expression cannot be evaluated unless the variables in the expression have already been instantiated in the knowledge base entry.

Variables are useful for two distinct purposes. The first is to receive and pass the individual values of an expression. This purpose was illustrated above in the example:

rule3:

if preferred-body=X then recommended-body=X. During the application of such a rule, the variable X becomes instantiated to the first value of the expression "preferred-body" when the rule premise is tested and it is found that the variable X can assume an arbitrary value. The first value of the variable X is later transferred to the expression "recommended-body" when the conclusion is noted. This testing of the premise and noting of the conclusion is iterated for all of the rest of the values of the expression "preferred-body."

The second purpose of a variable is to pass values to a function in the knowledge base. An example of this use is illustrated by the "Factorial Calculator" knowledge base in Appendix X to the specification. The factorial function "fact-N" has an informal argument N merely appended to the functor "fact-". Functions may also have strings of formal arguments such as "functor(Arg1, Arg2, Arg3)". The function variables are instantiated to argument values when the knowledge base is searched to find the value of the function with stated argument values. The argument variables match with and are instantiated to corresponding argument values, resulting in the rule or fact being applied to return the function value.

The factorial function "fact-N" is recursively defined by Rule 1 in the "Factorial Calculator" of Appendix X. Since the rules may be recursive, the variable bindings are local to the knowledge base entries, and are stored in a variable table for each rule or fact that is currently being tried.

The modifications to the FINDCHAR subroutine of FIG. 14 to perform a simple method of binding argument variables is shown in FIG. 27. The modified subroutine generally designated 310 has a step 311 to detect the presence of a capital letter after the initial matching of an expression. Once the capital letter is detected, in step 312 the expression pointer IC and the knowledge base pointer EC2 are used to search forward to the next respective parsing characters (blanks or commas) to obtain the variable and matching argument value. In step 313 the variable table is searched for the variable. If the variable is not already in the table and hence unbound, as determined in step 314, the variable is bound in step 315 to its corresponding argument value. Otherwise, in step 316, the argument value is compared with the variable's binding. If the values do not match, all of the current bindings must be released by clearing the variable table for the current knowledge base entry in step 317 since the function cannot be applied. Otherwise, searching may continue for more argument values and function variables.

It should be apparent that an expression should be searched for variables and the variables replaced with their bindings in the very first step of the SEEK subroutine 220 in FIG. 20. Also, when an expression is set to a value, the variables in the value should be replaced with current bindings. Thus, the first step in the NOTE subroutine 280 in FIG. 25 should search for variables in the value to be noted and replace them with their current bindings. To prevent the stack of variable tables from overflowing, all bindings for the current knowledge base entry should be released just prior to returning from the TRY subroutine 250 in FIG. 22.

The instantiation, substitution, and deinstantiation of variables during the testing of a rule premise is illustrated in FIG. 28. In step 351 of the modified TEST subroutine 350, SEEK is called to insure that a value and certainty factor for the expression is stored in the cache. Then in step 352 the value V is inspected to determine whether V is a variable.

If V is not a variable, then in step 353 the value V is transferred to V'. In step 354 the next value and certainty factor for the expression E is obtained from the cache. If the end of the cache is reached, as tested in step 355, the premise fails. The logical flag S is set false in step 356 and execution returns to the calling program. Otherwise, in step 357, the value for the expression is compared to V' and the certainty factor is compared to 20. If the value for the expression matches V' and the certainty factor is at least 20, then the premise succeeds. The logical flag S is set true is step 358 and execution returns to the calling program. Otherwise, the premise might succeed for some other value of the expression E, and execution jumps back to step 354 to get another value and certainty factor for the expression E from the cache.

If in step 352 the value V was found to be a variable, then in step 359 the current variable table is searched to determine whether the variable V has already been instantiated. If the variable V is not in the current variable table, then in step 360 the value and certainty factor for the expression E is obtained from the cache. Then in step 361 the certainty factor is compared to 20 to test whether the premise can be satisfied, with sufficient certainty, by binding the variable V to the value and certainty factor. If not, then in step 362 the next value and certainty factor for the expression E is obtained from the cache. If the end of the cache is reached, as tested in step 363, the premise fails. the logical flag S is set false in step 364 and execution returns to the calling program. Otherwise, execution jumps back to step 361 to test the certainty factor.

Once a certainty factor of at least 20 is found, in step 365 the variable V is instantiated by adding V to the current variable table along with the value and certainty factor. In step 366 the logical flag S is set true and execution returns to the calling program.

If in step 359 the variable V was already in the step variable table, the variable table is checked in step 367 to determine if the variable has been marked "RELEASE" in a fashion described below. If not, in step 368 the variable's binding, or value, is obtained from the variable table, and in step 369 the value is transferred to V'. Execution continues in step 354 in the fashion already described.

The variable V is always instantiated to just a single value of the expression E at any given time. So that attempts will be made to apply the current rule for all possible instantiations, and all possible combinations of instantiations if the rule includes multiple variables, just before returning from the INVOKE subroutine of FIG. 23, the current variable table is searched for the cache pointers to determine whether a variable has been instantiated to satisfy the premise of the current rule. If so, the last of such variables is marked "RELEASE" and execution jumps back to step 261 in FIG. 23 in an attempt to satisfy the rule using a different set of variable instantiations. This process of jumping back to try different variable instantiations is called "backtracking".

If in step 267 of FIG. 28 it is found that the variable V is marked "RELEASE", then in step 370 the old cache pointer for the variable v is obtained from the current variable table and the current cache pointer is replace by the old cache pointer. Thus, the old cache pointer stored in the variable table keeps track of where the last used value for the expression E was obtained from the cache. In step 371 the variable V is removed from the variable table so that the re-instantiation process may continue in step 362 in the same fashion as described above.

One of the major advantages of implementing BEST in PROLOG is that PROLOG inherently instantiates variables. PROLOG easily perform backtracking and inherently matches variables in a function argument list when the knowledge base is searched.

In a preferred PROLOG embodiment, the KBLOAD subroutine of FIG. 18 generates a single PROLOG fact for each knowledge base entry. Each PROLOG fact is a unique definition of the predicate "kbentry" and the corresponding knowledge base entry is the argument for the predicate. In other words, the PROLOG data

______________________________________kbentry([knowledge base entry 1]).kbentry([knowledge base entry 2]).kbentry([knowledge base entry 3])...______________________________________ The KBENTRY subroutine of FIG. 17 is then merely a PROLOG statement in the form of: kbentry(EXPR, L: KB). The variable L becomes instantiated to the label and the variable KB becomes instantiated to the rest of the current knowledge base entry. Matching of the knowledge base entry to a template is performed, for example, by the PROLOG statement: KB=if P then C

PROLOG also supports list processing by matching the arguments X and Z of the function "[X.vertline.Z]" to the first element and the list of the remaining elements, respectively, in a string of argument values. Thus, if the knowledge base entry is a fact in the form of:

factl: expr=[1, 2, 3, 4]. then execution of the PROLOG statement: kbentry(list, expr=[X.vertline.Z]) instantiates the variable X to the item 1 and instantiates the variable Z to the list [2, 3, 4]. Consider, for example, the knowledge base below: ______________________________________fact1: sum-of-[N] = N.rule1: if sum-of-REST = RSUM and FIRST + RSUM = SUM then sum-of-[FIRST.vertline.REST] = SUM.______________________________________ If the variable N is instantiated to a list of numbers and the knowledge base is searched for the value of the expression sum-of-[N], the variable SUM will become instatiated to the sum of the numbers in the list N which will become the declared value of the function sum-of-[N].

Another advantage of a PROLOG embodiment is that the knowledge engineer can write PROLOG subroutines. To aid the knowledge engineer, it is desirable for BEST to include a built-in proposition "prolog(EXPRESSION)" that, when found in the knowledge base and evaluated by the subroutine SEEK, passes the EXPRESSION to the PROLOG module (42 in FIG. 3) for execution as a PROLOG subroutine. Uninstantiated variables in the EXPRESSION serve as return parameters. It is also desirable to permit the PROLOG EXPRESSION to call directly at least some of the BEST subroutine modules (FIGS. 4 and 5) so that the PROLOG EXPRESSION may look up or note values for expressions in the cache.

A program listing of a PROLOG version of BEST is included in Appendix XI to the specification. As an illustration of the PROLOG backtracking feature, consider the highly simplified excerpt shown below:

______________________________________test(P and Q, CF) :- test(P, CF1), test(P, CF2), minimum(CF1, CF2, CF).test(E = V, CF) :- unless(sought(E), seek(E)), lookup(E = V cf CF because R).seek(E) :- assert(sought(E)), kbentry(E, KB), try(KB), !seek(E).try(L: if P then E = V cf CF1) :- test(P, CF2), CF is (CF1 * CF2) / 100, assert(lookup(E = V cf CF because L)), fail.try(L: question(E) = Q) :- write(Q), read(V), note(E = V cf 100 because yousaidso).______________________________________ Note that the first set of clauses for TRY finishes with a FAIL statement. This means that the first set of clauses can never succeed. Immediately after noting the conclusion, the first set of clauses fails causing backtracking into TEST to see if TEST can produce another way of binding the variable that makes the premise true. Suppose, for example, the cache contained the following two facts: best-color=red cf 70 because rule1. best-color=white cf 40 because rule2. That is, there is evidence that the best color should be red and there is also evidence (but less) that it should be white. Suppose the knowledge base contains the following rule: rule3: if best-color=C then recommended-color=C cf 80. That is, if there is a C such that there is evidence that C is the best color to have, then recommend C. Consider what happens if we now SEEK(RECOMMENDED-COLOR). The knowledge base is searched for entries relevant to RECOMMENDED-COLOR. Suppose that only rule 3 is found. It will be tried, thus: try(rule3: if best-color=X then recommended-color=X cf 80). This will cause BEST to test the premise of the rule: test(best-color=X, CF2). Since BEST-COLOR is already sought, this TEST is satisfied by a simple lookup to the cache. The first matching entry binds X to RED and CF2 to 70, so that NOTE is called: note(recommended-color=red cf 56 because rule3). Then BEST encounters the FAIL statement, backtracks into TEST, and picks up the second binding for X resulting in: note(recommended-color=white cf 42 because rule3). Then BEST encounters the FAIL a second time, and TEST also fails since there are no more matching cache entries. TRY fails back into SEEK which tries for another knowledge base entry. If there are more knowledge base entries for recommended-color, then they are TRYed; if not, the first clause of SEEK fails and the second `dummy` clause catches the failure.

The simplified mechanism as described above has some value but it is vastly improved by permitting expressions to be declared either single valued or multiple valued. The benefits of incorporating this distinction are both declarative and procedural. Declaratively, a single valued expression is one that can have only one value. If some particular value is certain that even if there is evidence for other values, those other values can be ignored. A multivalued expression may, of course, have many values. For a multi-valued expression, the fact that one value becomes certain is of limited interest. It is, however, the way the distinction is implemented that is of real value. BEST enables the knowledge engineer to write conventional procedures as well as embody heuristic knowledge.

To incorporate the single valued/multivalued distinction in the PROLOG code above, instead of simply FAILing at the end of TRY, the expression is tested to determine if it is now certain:

______________________________________try(L: if P then E = V cf CF1) :- test(P, CF2), CF is (CF1 * CF2) / 100, note(E = V cf CF because L), kbentry(E, multivalued(E)), lookup(E, ANYVALUE, 100, ANYREASON).______________________________________ The last two steps of this procedure succeed if E is singlevalued and definite. Since TRY succeeds, SEEK does NOT fail back into KBENTRY and so the SEEK is complete. Further features and advantages will become apparent to the reader familiar with PROLOG upon inspection of the PROLOG listing in Appendix XI to the specification.

In view of the above, a useful and versatile knowledge engineering tool has been described for building knowledge systems and running consultations. The knowledge engineering tool is easily mastered by people with little computer experience yet also provides advanced capabilities for the experienced knowledge engineer. The language of the knowledge base is transparent, universal, and extensible, and supports certainty factors, recursion, symbolic or universally quantified variables, and user-defined functions. The knowledge engineering tool is easily implemented in PROLOG to enable the knowledge engineer to write and add PROLOG subroutines to the knowledge base. ##SPC1##

Claims

1. A knowledge engineering tool comprising

a computer having memory for storing a knowledge base and memory for storing predetermined instructions including

means to augment and edit the knowledge base, and

means for interpreting the knowledge base to run an interactive consultation with the user to determine the value of at least one predetermined goal expression,

wherein said knowledge base includes

rules including premises having logical operations and corresponding conclusions concluding at least one value for said goal expression, said rules including recursive rules and rules having universally quantified variables,

propositions including arithmetic operations,

declarations defining whether selected expressions are single-valued or multiple-valued,

declarations defining questions to be asked to determine the values of selected expressions, and

declarations defining legal values of selected expessions, and

wherein said means for interpreting the knowledge base includes

means for determining the value of said goal expression including

means for invoking and chaining said rules to conclude a value for said goal expression,

means for evaluating the logical operations in the premises of the invoked rules,

means for evaluating said propositions,

means for instantiating and deinstantiating the universally quantified variables in said rules,

means for asking the user said questions to determine the values of their corresponding expressions,

means for receiving from the user and checking values of said corrsponding expressions against the corresponding declared legal values, and

means for explaining to the user why the user is asked for the values of said corresponding expressions,

means for conveying to the user said value of said goal expression,

means for explaining to the user how said value of said goal expression was determined, and

means for generating a record of said consultation including a list of the rules invoked during a test consultation.

2. A knowledge engineering tool comprising

a computer having memory for storing a knowledge base and memory for storing predetermined instructions including

means to augment and edit the knowledge base, and

means for interpreting the knowledge base to run an interactive consultation with the user to determine the value of at least one predetermined goal expression,

wherein said knowledge base includes

rules including premises having logical operations and corresponding conclusions concluding at least one value for said goal expression, said rules including rules having universally quantified variables,

certainty factors associated with said rules for encoding uncertain as well as certain knowledge,

propositions including arithmetic operations,

propositions responsive to whether a value of a predetermined expression has been determined with a predetermined degree of certainty, and

wherein said means for interpreting the knowledge base includes

means for determining the value of said goal expression including

means for invoking and chaining said rules to conclude a value for said goal expression,

means for evaluating the logical operations of the premises of the invoked rules and selecting certainty factors for the concluded values based on the certainty factors of the rules and the certainty factors of the premises,

means for updating the certainty factor of a value of an expression by combining certainty factors,

means for evaluating said propositions,

means for instantiating and deinstantiating the universally quantified variables in said rules, and

means for conveying to the user said value of said goal expression and its certainty factor.

3. The knowledge engineering tool as claimed in claim 2, wherein at least one of said universally quantified variables represents an argument of a function defined by one of said rules, and said means for instantiating and deinstantiating the universally quantified variables in said rules includes means for instantiating the function argument variable to a respective function argument value.

4. The knowledge engineering tool as claimed in claim 2, wherein at least one of said universally quantified variables is used in both the premise and conclusion of one of said rules, and said means for instantiating and deinstantiating instantiates the variable in the premise of the rule thereby enabling the value of the instantiated variable to be used in asserting the conclusion when the rule is applied.

5. The knowledge engineering tool as claimed in claim 2, wherein the same variables are used in more than one rule, and said means for instantiating and de-instantiating includes means for localizing variable bindings to the rules in which the variable bindings are applied.

6. A knowledge system comprising a computer having a memory, said memory storing a knowledge base including expressions and knowledge for determining values for said expressions, said memory including a cache porton for storing certain ones of said expressions and corresponding determined values for the stored expressions, said memory also storing a control procedure executable by said computer for interpreting the knowledge base to determine values for said expressions, said control procedure including

means for recognizing the occurrence of an expression in the knowledge base, means for searching the cache to determine whether the recognized expression is stored in the cache, means for obtaining from the cache the corresponding value of the recognized expression when the recognized expession is in the cache, and means for applying the knowledge in the knowledge base to determine a corresponding value for the recognized expression when the recognized expression is not in the cache and then storing the recognized expression and its corresponding value in the cache.

7. The knowledge system as claimed in claim 6, wherein said means for interpreting the knowledge base to determine a corresponding value includes means for determining a value indicating that the expression is unknown.

8. The knowledge system as claimed in claim 6, wherein said cache includes memory space for storing certainty factors corresponding to the value of said certain ones of said expressions, and said control procedure includes means for interpreting the knowledge base to determine a certainty factor for the value determined for the recognized expression when the recognized expression is not in the cache and then storing the determined certainty factor in the cache along with its corresponding expression and value.

9. The knowledge system as claimed in claim 6, wherein said cache includes memory space for storing reasons for the values of said certain ones of said expressions, and said control procedure includes means for interpreting the knowledge base to determine a reason for the value determined for the recognized expression when the recognized expression is not in the cache and then storing the determined reason in the cache along with its corresponding expression and value.

10. The knowledge system as claimed in claim 6, wherein said cache includes memory space for storing multiple values for some of said certain ones of said expressions declared in the knowledge base as multiple-valued, and wherein said means for applying the knowledge in the knowledge base determines all of the values for a multiple-valued expression and stores them together in the cache.

11. The knowledge system as claimed in claim 6, further including means for permitting a user to edit the contents of the cache.

12. The knowledge system as claimed in claim 6, further including means for permitting a user to interrupt execution of the control procedure at a desired point and to permit the user to set and reset the cache to a desired state.

13. The knowledge system as claimed in claim 1, wherein said control procedure includes means for clearing the cache of entries for a specified expession.

14. The knowledge system as claimed in claim 6, wherein said cache stores respective certainty factors for multiple values corresponding to certain expressions, and said control procedure includes means for clearing the cache of values having partial evidence when said means for applying the knowledge in the knowledge base determines a definite value for an expression.

15. The knowledge system as claimed in claim 6, wherein said cache stores certainty factors and reasons for values corresponding to certain expressions, and said means for applying said knowledge includes means for accumulating evidence from said knowledge base for determining values by combining certainty factors and reasons for values and storing the combined certainty factors and the combined reasons in the cache.

16. The knowledge system as claimed in claim 6, wherein said control procedure includes means for searching the knowledge base and finding the entries in the knowledge base which conclude a value for a specified expression.

17. The knowledge system as claimed in claim 6, wherein said control procedure includes means for searching the knowledge base and finding the entries in the knowledge base which conclude a specified value for a specified expession.

18. The knowledge system as claimed in claim 6, wherein said knowledge base includes condition-action rules and said means for applying said knowledge includes means for invoking and chaining said rules.

19. The knowledge system as claimed in claim 6, further comprising means for permitting a user to save the contents of the cache in a second memory, and to load the saved contents of the cache back into the cache.