Magnum One

home *** CD-ROM | disk | FTP | other *** search

/ Magnum One / Magnum One (Mid-American Digital) (Disc Manufacturing).iso / d26 / expert.arc / MANUAL.DOC < prev next >

Wrap

Text File | 1986-01-09 | 51.1 KB | 1,343 lines

TM T H E E X P E R T TM ApprovalWare for the Implementation of User-Specified Expert Systems Software and Manual by Stephen Walton 496 Hudson Street, Suite D15 New York, N.Y. 10014 Copyright (C) 1986 by Stephen Walton. All rights reserved. THE EXPERT is a trademark of Stephen Walton for his expert- system-implementation software. ApprovalWare is Stephen Walton's trademark for his user-supported software. Table of Contents Preface......................................................1 1. Introduction to THE EXPERT................................3 2. Running THE EXPERT........................................7 3. About Rulebases..........................................14 4. About Factbases..........................................18 5. Potential Enhancements...................................19 Preface THE EXPERT is software I once thought I would sell for a lot of money. It is possible that it's worth a lot of money. But there was a finite window in which it could be a commercial product, remain under my control, and make all that money through any channel available to me. That window -- if it ever really was there -- has closed. I will catalogue my errors of judgement and other sins some other time. I will NOT subject you to my course in non-marketing here. The software stands, and I feel good about standing by it. It works, it doesn't trip over its own ambitions, and it's clean. The gotchas are few and, better yet, documented below. Anyone who's been programming for a year or more (any language, any level) should be able to make THE EXPERT do pleasing things. It does assume that you have access to a text editor that produces ASCII files (non-document mode, you WordStar fanatics) and know how to use it. Beyond that, the software, this manual and the example files should be enough to get you going. ApprovalWare is my own stupid wrinkle on the "freeware" idea. I can't call this "freeware" because somebody else made an effort to trademark the "freeware" name. What ApprovalWare means is that if, after inspection, you approve of the software, you really should send me some money. It's a little like those stamp- collecting "approvals" you get stuck with whem you send away for the FREE 10,000 all-different citrus-fruits-of-many-lands stamps. But here there are no extras, and no building-up debt. You don't owe me anything until you decide you do. If you like the software and/or you find it useful, you can send me money. If you don't like it or find it useful, please pass it along to someone who may like it better (and you can still send me money, but I'll wonder about you). Your contribution, if you choose to make one, is a vote. It's a vote for the continued existence, conscientious support and energetic further development of this software. Your ongoing use of this software with no contribution is a vote for minimal or no support, for no enhancements and for the despair of a programmer who sincerely thought that ApprovalWare (or whatever alias you prefer) could be an equitable mode of software distribution. A couple of questions might arise quite naturally here: Q. Is THE EXPERT in its present form merely a stalking horse for a commercial product? A. Yes -- IF it generates some kind of commercial response. It would be a pleasure to discuss enhanced software, specific applications or classes in knowledge engineering. Custom work of every kind is possible. Every contributor-user will have a telephone number to use. THE EXPERT page 1 Q. Is THE EXPERT the precursor of a better ApprovalWare product? A. For sure -- if there's any approval out there. I've got a few notions for the next version, but the needs that appear in the field will take priority, if I'm kept informed. Every contributor-user will have a telephone number to use. Q. Will circulation of systems based on THE EXPERT be encouraged? A. Of course. If enough contributor-users get things going, their advertisements will be included in the ApprovalWare distribution of future releases of THE EXPERT. Every contributor-user will have a telephone number to use. Details are pending, but they will be limited to one screen each, I will take no responsibility for their content, AND I may censor or suppress them if I think they may offend general standards of good taste, or at my own judgement. Additionally, other tests may be imposed by public- domain software distribution channels and will not be my responsibility. Other means by which contributor-users can inform other contributor-users of available applications, such as a rudimentary newsletter, will also be explored. Quite frankly, I really don't know whether THE EXPERT in its present form is a complete product or just a hobbyists' curiosity (if that). I'm looking forward to finding out myself. Even if only hobbyists take an interest in it, that interest would please me, since the microcomputer hobby has been severely undervalued in the past few years and deserves more support than it currently gets. In the text that follows, "we" is used only as an editorial conceit. I am solely responsible for the contents. Where the software is concerned, an acknowledgement is owed to Jim Baen who graciously provided me with an early version of his proprietary QuickScreen package, which made it easy for me to do the popup menus. Finally, and whether you ever send a contribution or not, I thank you for going as far as this in examining THE EXPERT. I hope you will go farther, and find it worth your trouble. Stephen Walton New York City January 9, 1986 THE EXPERT page 2 1. Introduction to THE EXPERT What is knowledge? How do you think a problem through? How do experts do what they do? How many ideas can a person keep in mind at one time? These are practical as well as theoretical questions. Practical because there is a serious effort under way, in this country and elsewhere, to work out means of making computers perform more and more of the functions that we lump together as "thinking" when they are performed by humans. This effort is called Artificial Intelligence, nicknamed AI. A large portion of this effort is in the area of "expert systems," also known as "knowledge-based systems" -- the distinction between knowledge and mere data a significant one. The idea is that of putting useful knowledge, not just random facts, into a form in which a computer can successfully employ it to solve real problems as well as human experts do. Bodies of knowledge addressed so far have included medical diagnosis, analysis of mass spectrograms, petroleum geology and the recognition of natural speech. 1.1 What THE EXPERT Is THE EXPERT is software for the implementation of expert systems. It has the following features, which in combination we believe to be unique: 1. It is a framework on which any number of expert systems can be built, and is not itself a single such system. 2. It runs on microcomputers, and runs fast enough to use interactively. 3. The format it requires for a body of knowledge is very simple (this, in turn, imposes certain limitations which are discussed below). 4. Its results (conclusions and their justifications) can be retained for future examination or other use. 5. It works both forward and backward -- either deriving consequences from new data or testing hypotheses put to it as queries -- and these modes can be freely mixed by the user. 6. It can explain what it does -- why it is asking a particular question or how it arrived at a particular conclusion. 7. It allows you to change your input at any time, and THE EXPERT page 3 immediately re-evaluates all its conclusions. 8. Because it permits chained blocks of knowledge, it is not limited in the complexity of the subjects with which it can deal. 9. It does not require a hard disk or an unusually great amount of memory (128K RAM suffices). In AI parlance, THE EXPERT uses knowledge represented as "production systems." A production is a rule; it specifies that IF a list of conditions is met, THEN a list of results is true (sometimes a result is an action to be performed, rather than a new fact). A production system is a collection of productions, all presumably relating to a common subject or area of knowledge. Production systems are explained in detail in Patrick Henry Winston's textbook, Artificial Intelligence (Addison-Wesley, 1977). We will refer to them here as "rulebases," a term we consider slightly more congenial. 1.2 What THE EXPERT Can Do It would be impossible to enumerate here all the possible applications -- separate expert systems -- that can be done with THE EXPERT. We will, however, make so bold as to present the following list of application areas in which THE EXPERT might be used: Insurance underwriting Medical diagnosis Stock selection Psychiatric screening Computer system configuration Determinative mineralogy Taxation Pension-plan management Botanical and zoological taxonomy Software support Welfare eligibility Intelligent checklist for travel Sales-prospect analysis Hardware troubleshooting -- all types Software debugging Training -- all types Soil analysis Petroleum geology Insecticide or herbicide selection and application Fertilizer selection and application Engineering design Job scheduling Preventive maintenance Corporate distribution Weather forecasting Formal logic THE EXPERT page 4 Experiment planning Any application for which decision tables have been used Project feasibility Project management Elementary education Engineering materials selection Insurance policy selection Horserace handicapping Licensing checklists for: Securities brokers Insurance agents Real estate agents Air transport workers Selection of tax-advantaged investments Parliamentary procedure Yacht-racing rules Livestock management Contract drafting (selection of boilerplate clauses) Securities registration Home and car maintenance for the do-it-yourselfer Advertising media analysis With that little sales pitch behind us, this is also the appropriate place to note what, in general terms, THE EXPERT works with, and to note some of its limitations. THE EXPERT takes both rules and facts in the form of ASCII strings, so it is not, at present, suited for any direct manipulation of graphic information. A rule is a list of potential facts -- conditions -- and another list of facts -- results -- that become true when the list of conditions is met. Both conditions and results can be stated in either positive or negative form. Each list must have at least one member; there is no structural limitation on the number of members a list may have. The results obtained from THE EXPERT can be only as good as the rulebase used. At present, it is entirely possible to construct worse-than-useless rulebases. If it proves feasible, a rulebase-editing utility program will do some testing of logical consistence and overall quality of rulebases. For now, one should keep in mind that THE EXPERT does NOT eliminate thinking -- it merely requires that one do it right just once. The rationale behind the development of THE EXPERT has been simple. We considered the emergence of expert systems for microcomputers inevitable. Taking the highly successful example of VisiCalc, we felt it would be best to allow users to prepare their own knowledge bases, just as they prepare their own spreadsheet templates. We selected formats for rules and facts that would allow convenient, speedy operation. While much AI work has gone into means of mediating between short-term and long-term memory, we felt that the most useful, practicable product would be one that simply let the computer act as a much-enlarged short- term memory. And we tried to keep everything very simple. THE EXPERT page 5 The block diagram below shows the general organization of THE EXPERT. ____________ ___________ | | | | | Rulebase |---------| | User | | | | | |____________| | Inference | | | ___________ | Engine | | | | | | | ______|______ | | Factbase |---| | | | | | | | |-----| | | |___________| |___________| | | | | | | User | __|_________|_______________|_____ | Interface | | | | | | Files Manager |-----| | |__________________________________| |_____________| | | Diskette(s) and/or hard disk The Rulebase is the production system or collection of rules that determines what conclusions can be drawn and what hypotheses can be tested. The Factbase is the collection of facts that the user and THE EXPERT are currently working with. The Files Manager gets rulebases and factbases from disk and saves factbases back to disk. The Inference Engine does the real thinking -- it draws conclusions from facts given by the user, and also determines what questions must be asked in order to prove or disprove a user-supplied hypothesis. The User Interface mediates between the human user and both the Inference Engine and the Files Manager. It also provides menus and review functions. Everything about THE EXPERT owes a heavy debt to Artificial Intelligence research. But in our own context we prefer the term "Intelligence Amplification." That is what THE EXPERT is meant to be about. THE EXPERT page 6 2. Running THE EXPERT Although we feel that in many cases menu-driven software can be stifling, we deliberately set up THE EXPERT as a menu-driven system, and we still feel no regrets -- especially inasmuch as the menus are the pop-up type and do not destroy what's below them on the screen, while the seasoned user can do away with them entirely. This chapter will NOT be a rote transcription of an operating session, for two reasons. First, because there's enough online help available to keep you out of trouble. Second, because we reserve the right to let the software change faster than the documentation. Here we will try to develop the concepts involved in running THE EXPERT, and to provide you with a sense of the space in which you're operating. 2.1 The Object of the Game When you're running an application -- a rulebase -- under THE EXPERT, you are usually doing one of two things: 1. Presenting a set of facts to see what conclusions, if any, can be drawn from them. This is forward reasoning. 2. Posing a potential fact as a hypothesis, then affirming or denying other potential facts (with "don't-know" response as a valid option) until your original hypothesis can be either confirmed or disproved, or until it is clear that neither confirmation nor disproof is possible on the basis of your answers. This is backward reasoning. There is a third, "browsing" mode that is a combination of the two above but requires no specific discussion here. The first mode above allows you to either type in your facts (if you really want to) or, more likely, select them from a "pick list" of potential facts relevant to the particular rulebase. The second mode above is the one most often actually used, especially since it ties in with the possibility of a default or "preset" hypothesis or query written into the rulebase as a suggested goal to work toward. Let's restate some of this a little differently. A rulebase defines implication relations among a set of potential facts, or what we also call fact formats. When you work with THE EXPERT and a rulebase, you are assigning truth values to specific fact formats as you assert or deny them as facts in the particular case under consideration. (When you attach "don't know" to a fact format, you are assigning it a special truth value that means it is a roadblock that THE EXPERT must detour around, if it can; THE EXPERT page 7 actually, THE EXPERT has four possible truth values for each fact format: asserted, denied, user-doesn't-know and wholly- untested.) When you use forward reasoning, you are selected the fact formats to assert or deny, and letting the conclusions fall where they may. When you use backward reasoning, you have told THE EXPERT which fact format to use as a goal, and you are letting it select the other fact formats which, once asserted or denied, could lead according to the rulebase to an assertion or a denial for a the goal fact format. Backward reasoning is more likely to be used in practice; with a well-constructed rulebase, it can direct the user to the relevant issues with no time wasted on irrelevancies. Perhaps the "case" idea should be made clearer. A rulebase defines a finite subject area or universe of discourse -- a realm in which only so many propositions can be true or false. Each use of THE EXPERT (granting that one session can contain many uses) is the consideration of a single case in that universe of discourse -- one set of true and false propositions that is complete in that universe in that it is the "given facts" you have provided by your assertions or denials PLUS the full set of conclusions that can be drawn, according to the rulebase, from those given facts. Maybe we'd better break for a visual aid. 2.2 The Cockpit THE EXPERT is divided into seven separate operating sections, each with its own menu and its own functions. Because of the way you use cursor keys to get from one to the other, they can be imagined to have a specific spatial arrangement. Here's the diagram: THE EXPERT page 8 _____________________________________________________________ | | | Main Menu (housekeeping functions) | |_____________________________________________________________| ^ ^ ^ ^ ^ | | | | | v v v v v _____________________________________________________________ | | | | | | | | | | | | | Rules | Fact | Basic | Given | Derived | | | | | | | | Review | Formats | Input | Facts | Facts | | | | | | | | | | | | | | -> | <- -> | <- -> | <- -> | <- | |____________|____________|___________|___________|___________| ^ ^ ^ ^ ^ | | | | | v v v v v _____________________________________________________________ | | | Exit Menu | |_____________________________________________________________| This is what the sections of the control panel are about: Main Menu: Select and load a rulebase, select and load a factbase (a partial or complete case), save a factbase; go to any other section -- but to go to any section but the Exit Menu, you must load a rulebase first. Rules Review: Scan the rules of the (loaded) rulebase, one by one, in an intelligible format; move right (cursor key) to Fact Formats. Fact Formats: Scan the fact formats (potential facts) of the rulebase, with the option of asserting or denying the current one (the one last printed on the screen), or making a query (hypothesis) of it; move left or right (cursor keys) to Rules Review or Basic Input respectively. Basic Input: Elect to use the rulebase's preset query, if it has one, type in your own fact format for assertion, denial or query (if you really want to -- it must match a fact format in everything except case of alphabetic characters); move left or right (cursor keys) to Fact Formats or Given Facts respectively. Given Facts: Scan the facts that YOU have provided by assertion or denial, with the option of erasing or reversing the truth value of the current one (the one last printed on THE EXPERT page 9 the screen). This is the really neat thing that we stole from the spreadsheet people: when you erase or reverse a given fact, ALL derived facts are wiped out and ALL possible conclusions that can be drawn from the altered set of given facts are freshly drawn. Additionally, you can eliminate all your given facts, or all your previous don't-know responses so that you can retry a query that wound up blocked, or both. Move left or right (cursor keys) to Basic Input or Derived Facts respectively. Derived Facts: Scan the facts that THE EXPERT worked out for itself on the basis of your input, display the derivation of the current one (the one last printed on the screen) showing the rule involved and which of its conditions were obtained directly from you and which were themselves obtained by derivation from rules. Move left (cursor key) to Given Facts. THE EXPERT's menus all work the same way: Press the key indicated for the function you want OR select the current -- blinking -- function by pressing Enter. You change the current function by scanning through the menu with the PgUp and PgDn keys (not the up-arrow and down-arrow keys, because these have other specific purposes in some sections). 2.2 Outline of Operations THE EXPERT is a single program, EXPERT.EXE and is invoked from the DOS command line in the usual way; no options are used. The software includes no overlays, but it MUST find the file EX.MNU on the current drive or it will not run past the signon. The online help files EX0.HLP through EX7.HLP are optional, but should be kept available on the current drive for at least your first few sessions. After its signon material, THE EXPERT goes to the Main Menu section. From here, you select and load a rulebase to use. Then, most often, you use option 7 to go to the Basic Input section. Typically, a rulebase will contain a preset query -- a default hypothesis or suggested goal for backward-reasoning operations. Often, the preset query will be something as simple as "done," and THE EXPERT will attempt to find a route to the "done" goal-state. What happens when you reason backward on a query, either preset or your own? THE EXPERT searches the rulebase for a rule by means of which the query can be asserted or dell thed as a consequence or result, assuming all the conditions (which may be positive or negative) of the rule can be met. If no such rule can be found, you are told that your query is not deducible. If such a rule is found, THE EXPERT begins asking you about its conditions. THE EXPERT page 10 You are asked about the fact formats that are the conditions in the rule that could prove or disprove your query, and you can assert or deny each with a single keystroke. If the pattern of assertions and denials that you give (thus turning the fact formats in question into given facts for this case) matches the pattern of assertions and denials in the rule's conditions, then your query is either asserted or denied as a derived fact -- along with any other results that the particular rule calls for. If this were the whole story, it would be an uninteresting one. In fact, THE EXPERT lets you do more -- and less -- than just assert or deny a rule's conditions. It also allows you a don't-know response, which tags the fact format involved as a kind of roadblock to further reasoning. THE EXPERT then finds another way, if it can, to assert or deny your hypothesis. It looks for another rule that can be used to prove or disprove the query without using the fact format about which you professed your ignorance. If it can find no such rule, it tables your original query, and takes up the fact format that got the don't-know from you as a new, temporary goal. Once that goal is satisfied one way or another, it can go back to the original one. Another way to put it: When a don't-know answer from you blocks THE EXPERT's progress toward a goal, it finds another route to the original goal, OR takes the new unkown as an intermediate goal, "stacking" the original goal to return to when it can profitably do so. In the backward-reasoning interrogation, you have two more options. One is the "Why ask?" response. It makes THE EXPERT show you the rule it is trying to apply, then return the question. The other option is the "Stop asking" response, which tells THE EXPERT to drop a query entirely; this response will be most useful when you realize you've set the software on a useless and tedious quest, or when you are debugging a possibly shaky new rulebase. THE EXPERT will sometimes get stuck in trying to handle a query, and will tell you so. Do not confuse this situation with a bug in THE EXPERT! When the software "can go no further," it's because the information you have provided, combined with the rulebase, leaves no reasoning path to the goal. This can easily happen when you say "don't-know" about a crucial fact format that cannot be derived from any others (call it "base level" in that case) or for which no possible derivation path remains open. This is when the "Clear don't-knows" option in the Given Facts section becomes valuable, because it lets you retry the original query, posssibly reconsidering some of your don't-know responses. One other bit of behavior that may be unsettling if it takes you by surprise is rulebase chaining. Since not every sphere of human knowledge and expertise can be compressed into 100 or fewer if-then rules, it is sometimes necessary (or convenient) to THE EXPERT page 11 divide an application into "chunks" of up to 100 rules and make use of THE EXPERT's ability to shift from one rulebase to another on the fly. When this is happening, you will see derived facts such as "CHAIN 0 foo" or "CHAIN 1 bar" displayed, followed by a fair amount of disk activity. Details are in the section on rulebases below, but what's basically happening is that we're moving to FOO.RUL or BAR.RUL and making sure that any facts we had that are also relevant to the new rulebase are available to it. In the "CHAIN 1 bar" case, we will immediately take up the preset query in BAR.RUL, if it contains one. You get to save your work, if you wish. This is what the "Save factbase" option in the Main Menu is about -- your opportunity to document and perhaps reuse your work on the case you've been considering. You can save it to disk, in which case THE EXPERT will be able to reload it, or you can "save" it to your printer as documentation of your input and the rulebase's conclusions. In either case, you get to provide an identifying one-line header. 2.4 How to Crash We believe in keeping all the bad news, or potential bad news, centralized. This is where we keep a list of the hazards one might encounter in running THE EXPERT. If you come up with a new one, PLEASE let us know as quickly as possible, whether you are an enrolled contributor-user or not. You could be saving others some grief. Consequences could run from mild annoyance to loss of data. The ones we know about we hope to fix before any further software release. The list follows: * Trying to do ANYTHING with a write-protected diskette. THE EXPERT sometimes has to create files even when it doesn't seem to be doing any such thing, and such an attempt made with a write-protected diskette could lead to subsequent directory trashing. These days, write-protect tabs should be used on archival master copies ONLY. * Trying to run THE EXPERT when the EX.MNU file is not available on the current drive; produces a polite but unproductive exit. * Trying to load a badly terminated rulebase, one that has no rule terminator at the end of the last rule (we still do this from time to time) or contains blank lines after the last rule's terminator; the effect is that of a system halt that requires a reboot, but no damage should be done. * As above, for trying to load a factbase that ends wrong. * Calling for printer output when no printer is available (it must be connected to the computer, turned on and on THE EXPERT page 12 line). Could cause program abort and loss of your work from this session. * Specifying a nonexistent disk drive for any operation. Could cause program abort and loss of your work from this session. * Trying to run THE EXPERT on a computer that is not CLOSELY compatible with the IBM PC/XT/AT family. * Trying to run with too little memory. If you are still using DOS 1.x, 128K should suffice. For DOS 2.x or 3.x, 192K or better is recommended or some larger rulebases will be unusable. Please be alert. We try to be user-friendly, but this isn't kindergarten. THE EXPERT page 13 3. About Rulebases The rulebase is the heart of a useful application under THE EXPERT. This chapter is where we get to tell you what they look like, and a little bit about how they should work. 3.1 The Format A rulebase is an ASCII file (beware of word processors in document mode) of type .RUL. It should not contain any extra blank lines at beginning or end. Every rulebase has these elements in common: A three-line header, presumably used to identify the rulebase in more detail than a filename can provide to the user. Zero or more rules, the form of which is detailed below. We will omit from consideration here those rulebases containing no rules, since they are less useful than others. The header is immediately followed by the first rule. A rule has this format: One or more condition lines A separator line, which must begin with the character ">" and may include a comment following that character One or more result lines An end-of-rule line, which must begin with the character "#" and may include a comment following that character. The condition lines determine the facts that must be asserted if the rule is to have effect. They may begin with a tilde ("~") indicating that the negation of the remainder of the line is the condition in question (double negatives are no more proper here than in conversation with educated speakers of English). Result lines follow the same constraints. Surprise! Surprise! Condition and result lines had better not begin with the characters ">" and "#"! Here's a very primitive sample: THE EXPERT page 14 Rain and sidewalks [header line] [header line] [header line] It's going to rain [single condition line] > [condition-result separator] The sidewalks will get wet [single result line] # [rule terminator] The notes in square brackets above are not part of the rulebase. 3.2 Your Options We tried to keep the rulebase format pretty free. You have a bundle of options: 1. Number of rules -- as many as 100 in the current version 2. Number of distinct line formats -- as many as 150 in the current version 3. Length of a single rule -- at least one condition line and one result line; but as many of each as you need 4. Negation -- is specified with a tilde ("~") at the beginning of a condition or result line 5. Length of a single condition or result line -- consider anything over 200 characters potential trouble 6. Comments -- may occur on conditions-results separator line and rule-terminator line; may be as long as 250 characters; will not be read in or be accessible to rulebase user. 3.3 PRESET Queries A special convention applies when you want to nudge your user toward a specific query when using your rulebase. When the third line of the rulebase header begins with the word PRESET, all in capital letters, and the rest of the line is a valid fact format in the rulebase (it's up to you to ensure this), then that fact format is the preset or default query for the rulebase. This is the query that will automatically be used as a goal when the user calls for "preset query" from the Basic Input menu, OR when the rulebase is accessed via chaining mode 1 (see below). 3.4 CHAINed Rulebases THE EXPERT cannot hold an unlimited number or rules or facts THE EXPERT page 15 at one time. Some applications, in all the glory of their detail, may simply be too large to swallow all at once. But essentially any complex subject can be handled adequately by means of chained or cascaded rulebases. There is no reason why one result within Rulebase A cannot be a command (or request) to next use Rulebase B. The sequence of rulebases thus invoked, of course, can as well be a branching one as a sequential one. The way you make one rulebase use another is by including as a result of a rule a format like this: CHAIN n xxxxxxx where "CHAIN" is the chaining command, n is either 0 or 1 for the chaining mode, and xxxxxxxx is the name (without ".RUL") of the rulebase to be used next -- which much be available on the drive specified for rulebases (the current drive if no explicit specification has been given). When a CHAIN result is encountered, THE EXPERT completes any reasoning that it was doing, saves the current factbase to disk, loads the rulebase named in the CHAIN result, and reloads the saved factbase. If the chaining mode was 0, the chaining action is now complete. If the chaining mode was 1, THE EXPERT now takes up the new rulebase's preset query (if any) as a new goal. With either mode, the reason for saving a reloading the factbase is to make it possible for the two rulebases to share any facts whose formats appear in both rulebases. Try out, and then type out the rulebase MENU.RUL, which uses mode-1 chaining to branch into other rulebases. 3.5 Rulebase Organization and Logic A lengthy topic that we cannot treat fully here. We hope in the future to provide a powerful assistant, if not full automation, for rulebase creation. Below are some of the important basics for creating your own rulebases with your favorite text editor. Consistency: THE EXPERT will effectively not load a blatantly inconsistent rulebase exactly as given. If two rules have exactly the same conditions but yield opposite results, only one (the one occurring later in the rulebase) will be used. Subtler forms of inconsistency are not trapped by the software, though, and are the rulebase maker's responsibility. Start your rulebase in the simplest possible form, let it grow more complex THE EXPERT page 16 only gradually, and test it thoroughly at each stage. Binary explosions: Because THE EXPERT's if-then rules are and-linked only (you can't say "If A or B then C"), you can easily get into a binary explosion of rules when many different combinations of conditions can produce a single result. One way to prevent this is the judicious use of intermediate goals. In the getting-dressed-for-the-day and will-drafting example rulebases, you will see fact formats of the form "such-and-such has been accounted for." These are intermediates and "has been accounted for" is our code to the user that a don't-know response is appropriate when the item comes up in a backward-reasoning interrogation. A strategy: For most applications, it makes sense to start with a goal in mind, then to write rules that will make it possible to reach that goal in specific examples. Remember that real-world domains will usually NOT involve an exhaustive set of combinations of conditions. E.g., if you are working on bird indentification, it is readily apparent that there is not a real bird for every conceivable combination of size, plumage, beak shape, etc. THE EXPERT page 17 4. About Factbases You've worked with some rulebases and some factbases by now. It's time we told you what a factbase is supposed to look like in the raw. 4.1 The Factbase Format A factbase (filetype .FCT) is an ASCII file and always contains these elements: An identifying header line Zero or more two-line fact statements, consisting of: A fact assertion or denial (starting with a tilde if a denial) A provenance value -- "Given" if the fact was given by the user, otherwise "Rule n" where n is the number of the rule used to derive the fact. Except for the header line, a factbase may contain comments only within the provenance lines (but after the provenance values). 4.2 Editing Cautions Since a factbase is an ASCII file, you can use your text editor on it freely. That may not be a good idea. A factbase can be thrown fatally out of synchronization with THE EXPERT's fact-reading routine [present version] if you eliminate the header line or make it more than one line, or if you drop either an assertion/denial line or a provenance-value line. We urge you to leave your factbases alone. TYPE and read them, certainly; print them out; amaze your friends and confound your enemies with them. But DON'T edit them. THE EXPERT page 18 5. Potential Enhancements This is the speculative section. We have some possible improvements in mind for THE EXPERT, and would like to know which ones you would find valuable. Please vote for your favorites. Or-linked rules: This would allow THE EXPERT to use rules of the form If A or B or ... then X and Y and ... This could make rulebase creation simpler by reducing or eliminating the need for intermediate (dummy) goals. Default don't-knows: Where those intermediate (dummy) goals are to be, they might be flagged in the rulebase in such a way that they are always automatically set to the don't-know state when the rulebase is loaded. This could make some operations seem smoother to the rulebase user -- or could make them rather mystifying. Confidence levels: An option for fuzzy or probabilistic reasoning, in which every rule may be given a trustworthiness value from 0.00 to 1.00, and every fact you assert or deny may be given its own confidence level on the same scale. The confidence level of a derived fact is then the product of the values for the facts on which it is based and the rule used, with a cutoff value, a minimum level that much be reached if the derivation is to be complete. This level could be set by the user or by the rulebase writer. Variables: THE EXPERT is an idiot savant inasmuch as it cannot do any arithmetic or even numeric comparisons. We might provide a mechanism by which variable names could be used within fact formats, for which the user would be asked for values when those formats occur in reasoning. These values could then be used in numeric comparisons, if not for more complicated arithmetic. Multiple cases and relations between cases: We sidetracked ourselves for a while with a scheme by means of which more than one case at a time could be handled, along with situations in which one case could affect general conditions for all. We abandoned it as irrelevant. If you think otherwise, tell us why. More action: By chaining from rulebase to rulebase, THE EXPERT shows that it can DO things, not just draw conclusions, and thus in some degree approaches the idea of an intelligent "software agent." What other things might THE EXPERT do that could be useful? A "shell" command that would permit loading and execution of other programs, without losing one's place within THE EXPERT, is a possibility. Support software: The Rulebase Editor that (embarrassingly) is mentioned in THE EXPERT's Exit Menu is on hold because it's worse than useless unless it's really smart. We are considering THE EXPERT page 19 other approaches, from a fairly simple logic checker for rulebases to a full-blown knowledge extractor, and do not yet know what will prove feasible. Got your own wishlist? Please share it with us! THE EXPERT page 20