SPECIALSECIKM!

RULE-BASEDSYSTEMS

Rule-based systemsautomate problem-solving know-how, provide a meansfor

capturing and refining human expertise, and are proving to be commercially
viable.

FREDERICK HAYES-ROTH

Rule-based systems (RBSs)constitute the best currently means of building automated experts in application
available means for codifying the problem-solving areas where job excellence requires consistent reason-
know-how of human experts. Experts tend to express ing and practical experience. Table I lists some applica-
most of their problem-solving techniques in terms of a tion areas currently addressed by RBS technology.
set of situation-action rules, and this suggests that RBSs The hallmark of these systems is the way they repre-
should be the method of choice for building knowledge- sent knowledge about plausible inferences and pre-
intensive expert systems. Although many different ferred problem-solving tactics. Typically, both sorts of
techniques have emerged for organizing collections of know-how are represented as conditional rules. Figures
rules into automated experts, all RBSs share certain key I and 2 illustrate rules that apply in a variety of appli-
properties: cations and that employ some of the basic syntactic
conventions of RBSs.
1. They incorporate practical human knowledge in
We can define RBSs as modularized know-how systems,
conditional if-then rules,
where know-how is practical problem-solving knowl-
2. their skill increases at a rate proportional to the
edge. Such knowledge consists of various kinds of infor-
enlargement of their knowledge bases,
mation, including
3. they can solve a wide range of possibly complex
problems by selecting relevant rules and then com- 1. specific inferences that follow from specific obser-
bining the results in appropriate ways, vations;
4. they adaptively determine the best sequence of 2. abstractions, generalizations, and categorizations of
rules to execute, and given data;
5. they explain their conclusions by retracing their 3. necessary and sufficient conditions for achieving
actual lines of reasoning and translating the logic of some goal;
each rule employed into natural language. 4. likeliest places to look for relevant information;
RBSs address a need for capturing, representing, stor- 5. preferred strategies for eliminating uncertainty or
ing, distributing, reasoning about, and applying human minimizing other risks;
knowledge electronically. They provide a practical 6. likely consequences of hypothetical situations;
0 1985 ACM oool-0782/85/0900-0921 750 7. probable causes of symptoms.

September 1985 Volume 28 Number 9 Communications of the ACM 921

 Special Section

TABLE I. Appliiations of RBBs that RBSs can directly incorporate rules that emulate
the effective special-case reasoning characteristic of
highly experienced professionals. General-purpose de-
Equipment maintenance Diagnosing faults and ductive schemes do not emulate experts and therefore
recommending repairs lack the efficiency necessary for solving complex prac-
Component selection Eliciting requirements and tical problems. Because each rule in an RBS approxi-
matching parts from an
electronics catalog
mates an independent nugget of know-how, these sys-
Computer operation Analyzing requirements, and tems have two characteristic features: First, existing
selecting and operating knowledge can be refined, and new knowledge added,
software for incremental increases in system performance. Sec-
Product configuration Eliciting preferences and ond, systems are able to explain their reasoning, mak-
identifying parts that satisfy ing their logic practically transparent and allowing
constraints them to satisfy the widely recognized need for under-
Troubleshooting Analyzing situatiis, suggesting standability in computer systems.
treatments, and prescribing By incorporating know-how acquired in an incre-
preventative measures mental and transparent manner, RBSs open up key
Process control Spottlng problematic data and
remedying inegufarfties
computing applications not readily addressable by al-
Quality assurance Assessing tasks, proposing ternative techniques. Some of these applications are in
practices, and enforcing areas where the supply of quality human workers is
requirements insufficient. There are a number of possible reasons for
this: The skill level among trained workers may not be
consistently high enough, experts may perform at a
level far beyond that of average workers, or conven-
Today’s RBS technology provides the first practical tional means of training and automation may fail to
methodology and notation for developing systems capa- produce adequate performance. Automating expertise
ble of knowledge-intensive performance. Although arti- in specialized tasks generally requires a few hundred to
ficial intelligence (AI) researchers have developed sev- a few thousand heuristic rules. With existing technol-
eral alternabves, only the RBS approach consistently ogy, this makes good economic sense in hundreds of
produces expert problem solvers. This reflects a feature application areas.
of the current state of the art. in automatic reasoning- Of course, in spite of such relevance and potential,

If the plaintiff did receive an eye injury

and there was just one eye that was injured
and the treatment for the eye did require surgery
and the recovery from the injury was almost complete
,and visual acuity was slightly reduced by the injury
and the condition is fixed,
increase the injury trauma factor by $10,000.
If the plaintiff's injury did cause
(a temporary disability of an important function)
and the plaintiff's doctors were not certain about
the disability being temporary
and the plaintiff's recovery was almost complete
and the condition is fixed,
increase the fear factor by $1,000 per day.
If the plaintiff did not wear glasses before the injury
and the plaintiff's injury does require
(the plaintiff to wear glasses),
increase the faculty loss factor by $7,500
and increase the inconvenience factor by $1,500.

The ROSIE programming system, developed at RAND, provides and actions. Each rule expresses an independent chunk of
a stylized English-like syntax for expressing conditions know-how.

flGURE 1. The Representation of Legal Heutistics for Product Liabilii in a ROBIEProgram

922 Communicatiom of the ACM September 1985 Volume 28 Number 9

 Special Section

Rule408:
C is a car.
If: the pattern observed by attaching an oscilloscope
to the charging circuit of the'car C is
fluctuating.arches, and
the alternator of the car C responds properly to
different loads,
Then: there is strongly suggestive evidence <.Q> that
the cause of the problem with the car C is
voltage.regulator.bad.
Rule428:
C is a car.
If: the pattern obtained by attaching an oscilloscope to the
charging circuit of the car C is straight.line, and
the result pulling out the field connector is no.flash, and
the field connector has does not have a voltage, and
the input of the voltage regulator does not have a voltage, and
the dashboard lights do not glow when their
ground circuit is completed, and
the fusable link is getting voltage, and
the fusable link is not conducting power,
Then: it is definite <l.O> that the cause of the problem with the car C
is fusable.link.bad.

As cars incorporate more electronic subsystems, they become StrUCtUral framework for organizing and applying thousands of
more difficult for the average technician to repair. rules. General Motors plans to aid its service technicians with
Teknowledge’s S.1 expert-system building tool provides a several large-scale RBSs.

FIGURE2. The Representation of Automotive Troubleshooting Rules in an S.l Program

RBS technology does have its shortcomings. As a new The technology for building RBSs has progressed sig-
technology, it will require years of perfecting and fine nificantly in the last 10 years, as many people have
tuning. Proposed applications must be assessedcare- analyzed the technology and assessedits relevance for
fully for feasibility and deployability. Only a small frac- a variety of tasks. Today, we can see that rule-oriented
tion of potential applications can be addressed today components are becoming central in many advanced
with off-the-shelf products. Nevertheless, RBSs consti- computing applications.
tute the best means available for building expert sys-
tems that incorporate large amounts of judgmental, AN OVERVIEW OF RBSs
heuristic, experiential know-how. Informal surveys in- Roughly speaking, an RBS consists of a knowledge base
dicate that approximately 50 percent of the Fortune 500 and an inference engine (see Figure 3). The knowledge
companies are investing in RBSs, and that about 10 base contains rules and facts. Rules always express a
percent have applications under development. conditional, with an antecedent and a consequent com-
The operating concept of RBSs differs radically from ponent. The interpretation of a rule is that if the ante-
von Neumann architectures. Intelligent problem solv- cedent can be satisfied the consequent can too. When
ing with RBSs involves an iterative cycle of (1) identify- the consequent defines an action, the effect of satisfying
ing from experience the heuristic rules that bear on a the antecedent is to schedule the action for execution.
problem at hand, and (2) applying one of those rules to When the consequent defines a conclusion, the effect is
solve or simplify the problem. The technology for to infer the conclusion.
building RBSs supports this cycle by providing a dy- Because the behavior of all RBSs derives from this
namic working memory for partial results, a device to simple regimen, their rules will always specify the ac-
identify relevant rules, and selective means for apply- tual behavior of the system when particular problem-
ing desirable rules. Many people conjecture that hu- solving data are entered. In so doing, rules perform a
man problem-solving activity follows the RBS model. variety of distinctive functions:
Whether or not that proves true, human experts gener-
ally find it easy to express methods for solving prob- 1. They define a parallel decomposition of state transi-
lems in their application areas by using a rule formula- tion behavior, thereby inducing a parallel decompo-
tion. sition of the overall system state that simplifies au-

September 1985 Volume 28 Number 9 Communications of the ACM 923

SpecialSection

Knowledge
base

Inputs .

A simple BE6 consists of.stomge and processingetements, patterns of data that can arise in working memory. The sys-
whii are aften referred to resgectivety as tha knowtedge tem represents data in terms of relations, propositions, or
base and the inference engine. The basic oycte of an BBS equivalent logical expressions. Facts define static, true prop-
consists of a seleot phase and an execute *aso During the ositiis. In contrast to conventional data-processing sys-
execute phase, the system interprets the s&o&d rule to tems, most RBSs distribute their logic over numerous inde-
draw inferences that alter the system’8 dynamic memory. pendent condition-act&n rules, monitor dynamic results for
System storage inclmtes components for Wig-term static trtggering patterns of data, determine their sequential behav-
data and short-term dynamic: data. The fonptann store, ior’by selecting their next activity from a set of candidate-
whii is ths krmwtadga base, contains rules and facts.Rules triggered rules, and store their intermediate results exclu-
specify actions the syetem should initiate when certain trig- sively in a global working memory.
gering conditions occur. These conditions deEne important

FIGURE3. The BasicFeaturesof an RBS

diting and explanation. Every result can thus be evant situations and corresponding actions. The lan-
traced to its antecedent data and intermediate rule- guage employed for these descriptions imposes a con-
based inferences. ceptual framework on the problem and its solution. The
2. They can simulate deduction and reasoning by ex- rules may be precise or gross; the intermediate partial
pressing logical relationships (conditionals) and def- solutions abstract or detailed. Efforts to solve the prob-
initional equivalences. lem may proceed top-down, outside-in, bottom-up, or
3. They can simulate subjective perception by relating in some other way. The meaning, importance, and con-
signal data to higher level pattern classes. tribution of each rule depend on its effectiveness as a
4. They can simulate subjective decision making by contributor within the entire set of rules available for
using conditional rules to express heuristics. solving a problem.
Facts, the other kind of data in a knowledge base,
Several key techniques for organizing RBSs have express assertions about properties, relations, proposi-
emerged. Rules can be used to express deductive tions, etc. In contrast to rules, which the RBS interprets
knowledge, :such as logical relationships, and thereby to as imperatives, facts are usually static and inactive-
support inference, verification, or evaluation tasks. implicitly, a fact is silent regarding the pragmatic value
Conversely, rules can be used to express goal-oriented and dynamic utilization of its knowledge. Thus, al-
knowledge that an RBS can .apply in seeking problem though in many contexts facts and rules are logically
solutions and cite in justifying its own goal-seeking be- interchangeable, in the context of RBSs they are quite
havior. Finally, rules can be used to express causal distinct.
relationships, which an RBS can use to answer “what In addition to its static memory for facts and rules, an
if’ questions, or to determine possible causes for speci- RBS uses a working memory to store temporary asser-
fied events. tions. These assertions record earlier rule-based infer-
An RBS can only solve problems if it incorporates ences. We can describe the contents of working mem-
rules that use symbolic descriptions to characterize rel- ory as problem-solving state information. Ordinarily, the

Communications of the ACM September1985 Volume 28 Number 9

 Special Section

data in working memory adhere to the syntactic con- 4. explanations of results, lines of reasoning, and
ventions of facts. Temporary assertions thus correspond questions asked;
to dynamic facts. 6. intelligibly encoded beliefs and problem-solving
The computing environment for rule interpretation techniques;
consists of current facts and the inference engine itself. 6. inference chains assembled dynamically by built-in
Together, these provide a context for interpreting the control procedures that can often perform efficient
current state, understanding what the rules mean, and searches.
applying relevant rules appropriately. Evidence of this
Given this wide range of important applications, it
implicit frame of reference can be found in Figures 1
seems probable that the role of RBSs in program devel-
and 2. The legal rules specify the changes that are to be
opment will be expanding.
made to various “factors” under various conditions, and
the auto repair rules draw conclusions about the causes
of problems. Of course, these rules are not universally THE RULE AS AN OBJECT
valid. Each depends on many unstated assumptions in We speak of a rule as a relatively independent piece or
its particular frame of reference. The validity of these chunk of know-how. Psychologists have for some time
rules depends critically on their being interpreted in emphasized the importance of chunks as elementary
the right context. Thus, RBSs cannot obviate all the patterns in perception and thinking. Chunks are a dis-
concerns of conventional computer programming (e.g., tinctly subjective psychological phenomenon: They re-
state sequences and variable scoping) because someone flect the learned, appropriate, effective distinctions that
must still ensure that an RBS applies rules appropri- people use to make sophisticated high-level decisions.
ately and in meaningful contexts. Many people mistak- A rule can serve as just such a chunk of problem-
enly assume that RBSs can turn unstructured heaps of solving know-how.
universally valid, independent rules into effective prob- As used by most RBSs, rules specify chunks of ana-
lem solvers. That is a serious misinterpretation of cur- lytic problem-solving knowledge. A rule is a datum em-
rent technology. Furthermore, rule writers must con- ployed by an inference engine to infer a solution to its
sider the rule interpretation environment to ensure that goal problem. Thus, a rule writer who expresses know-
a rule or its applications can be translated into appro- how in rule format is offering one possible path to re-
priate natural language. ducing a goal to subgoals, to drawing a plausible infer-
The basic function of an RBS is to produce results. ence from plausible data, or to transforming an expres-
The primary output may be a problem solution, the sion. This information about the rule typically com-
answer to a question, or an analysis of some data. prises its familiar if-then components. However, as the
Whatever the case, an RBS employs several key pro- number of rules in an RBS grows, a need arises for rule
cessesin determining its overall activity. A “world” components that can support multiple functions (see
manager maintains information in working memory, Figure 4) and thereby extend and maintain the knowl-
and a built-in control procedure defines the basic high- edge base. For this reason, many additional facets or
level loop; if the built-in control provides for pro- attributes are introduced to represent data about a
grammable specialized control, an additional process rule’s analytic knowledge and its preferred manner
manages branches to and returns from special control of use.
blocks.
RBS ARCHITECTURE
An RBS is generally a complete computing system,
THE RBS NICHE IN COMPUTING
which is to say that it can produce an output by apply-
RBSs address a number of shortcomings in conven-
ing memory and processing to an input. An architec-
tional programming technology, among them
tural inventory of RBS technology would include the
1. the nonspecifiability of programs, following four basic elements:
2. the rapid changes in principles of operation that
Rules. From an architectural perspective, rules are
can arise during development,
data that generally conform to highly specialized gram-
3. the lack of user/expert participation in operations
mars capable of using symbolic expressions to define
specification,
conditions and actions. Current systems differ primarily
4. the lack of experimental development for computer.
in the generality and notational convenience their sym-
based competence, and
bologies support.
5. the lack of expertise in exploiting computer capa-
bilities. Interpreters. The rule interpreter matches a rule com-
ponent to working memory data. Generally this re-
Among the features that allow them to do this are
quires pattern matching to find constants in working
1. modular know-how; memory that match identical constants or unbound
2. knowledge bases for storing rules and facts that di- variables in rule patterns. Existing systems differ pri-
rectly determine decisions; marily in the methods they use to simplify rule defini-
3. the capacity for incremental development with tion and pattern matching. The action of the rule is
steady performance improvements; produced by another part of the rule interpreter. Ac-

September 1985 Volume 28 Number 9 Communications of the ACM 925

Special Section

ljuperstructure Contained in rule sets

rI
Conditional IF <Antecedent> THEN <Consequent>
Date
Data Author . . .
Uses

Translations

Control

Rules can ccxltain more information than can be found in a was actually being used to make an inference. Typically.
simple if-then conditional. The antecedent and consequent each rule exists in several alternative representatiis or
specify data sufficient for inferring a conclusion or performing translations that suit different purposes. For instance, one
another action, while other parts of a rule serve additional machineoriented translation might serve tne need for hgh-
important ra4e.s.For instance, most large RBSs benefit from performance at run time, another human+riented form might
hierarcMal structuring, whereby individual rules can belong use English to support pubtiition and explanation, a third
to one or more higher order cottections. These rule sets could exploit terseness to make reading and editing easier,
aggregate and differentiate rules according to their functii and other facets of the rule structure might determine how
within the system. Thus, an RHS is capabte of ignoring rule the inference engine should treat the rule. The system may
sets that might be irrelevant to a particular probfem at a need to trace rule evaluations and applications, justify a
particular time. For instance, such data as who wrote a rule rule’s relevance, or even selectively ignore it under certain
and when would be retevant to testing, evaluation, exptana- conditions.
tion, and maintenance, but woukl be irrelevant when the rule

flGURE 4. The Organizationof a Ruleinto Componentsto SupportMultipleFunctions

tions generally fall into one of two categories: changes that matches its antecedent component. Some RBSs al-
to working memory or changes to external actions low rules to fire repeatedly as long as the working data
like I/O. still match the rule, but most process a specific working
memory data configuration only once for each rule. In a
Translations. Nearly all RBSs allow for multiple repre- back-chaining system, the RBS begins with a goal and
sentations of rules-one representation might be for successively examines any rules with matching conse-
data entry, .another for interpretation, and another for quent components. These candidate rules are consid-
explanations. Typically, all rules are maintained in one ered one at a time. The unmet conditions of the ante-
preferred representation and translated as needed for cedent are extracted from each plausibly applicable
other purposes. rule, and these conditions are in turn defined as new
Explanations. The hallmark of RBSs has been their goals. The back-chaining control procedure then shifts
ability to explain their conclusions. Explanations have attention recursively toward the new goal. The effort
been generated by translating the rules that contributed terminates when the top goal has been reduced to a set
to a decisio:n into natural language. This requires that a of satisfied subgoals.
history of working memory changes and their causes be From the point of view of computer architecture, two
kept that can be searched as needed for explanations. kernel facilities distinguish RBSs from conventional
systems. First, RBSs make heavy use of pattern match-
Although RBSs have been organized in a variety of ing between rule components and working memory.
ways, they all share a basic configuration-they are Second, they quickly identify rules that become rele-
sets of decisions about what meaning to give rules, and vant as working memory changes. This means that
how and when to interpret them. Two organizations there must be a way to access rules by pattern-matched
are most common: stimulus-driven or forward-chaining values. Most RBSs meet this need with software, al-
systems, and goal-directed or back-chaining systems. In a though some current hardware efforts are attempting to
forward-chaining system, a rule is triggered when improve performance for these tasks. Figure 5 shows a
changes in ,working memory data produce a situation representative sophisticated RBS.

926 Communications of the ACM September 1985 Volume 28 Number 9

 Special Section

IMPACTS AND APPLICATIONS and problem features explicitly, but can depend on the
OF RBS TECHNOLOGY RBS to automate almost everything else necessary for
solving the problem.
Evolutionary System Development We should anticipate that the complementary
RBSs have proved invaluable as a practical means for strengths of conventional programming and RBSs will
evolving poorly understood knowledge into a coherent motivate research efforts to bring the two technologies
knowledge base. Although today’s RBSs are no substi- together so that applications will be able to exploit the
tute for a full range of mature data-processing (DP) advantages of both.
application-building technology, they do offer a number
of unique advantages that are missing from the conven- THE CONCEPTUAL EVOLUTION OF RBSs
tional DP tool kit. We should anticipate that the essen- The RBS of today incorporates many influences and has
tial ingredients of RBSs will be imported into DP tech- partaken of many related technological developments
nology, as the technology matures, to assist in rapid (see Figure 6). Its essence derives from the production
p.rototyping, improving extensibility, and enhancing systemmodel used in automaton theory and psychol-
software maintenance and support. ogy. The basic model was the stimulus-response associ-
ation presumed by some to underlie all animal behav-
Searches ior. In a similar manner, computing theorists have
The focus on knowledge in applied AI systems repre- sometimes found it convenient to describe all computa-
sents a reaction to the unsuccessful attempt to solve tional behavior in terms of state transition tables
important problems using general-purpose or weak that define rules for moving between states. Oliver
methods. As the importance of knowledge became Selfridge’s early model, Pandemonium, viewed human
clear, many AI researchers became knowledge engineers. signal interpretation activity in terms of the actions of
These individuals set themselves the task of picking independent pattern-action modules called “demons.”
high-value problems with symbolic solutions, identify- Each demon listened to the “shouts” of subordinates
ing corresponding human experts, and debriefing these that were able to recognize constituent features that
experts to find out what they knew. were necessary for the demon’s higher order percept.
In many cases, expertise is the ability that some peo- After hearing all necessary inputs, the demon shouted
ple have to use shortcuts and labor-saving techniques its own message,thus indicating the perception of a
that less experienced persons would not know about. higher level pattern.
Experts are thus able to reduce the equivalent of a Many researchers have gravitated toward rule-based
large search space for a general problem-solving pro- representations of knowledge for two other reasons.
gram to that of the small search space of a specialized, First, rules seem like a natural way to express the
knowledge-intensive program. So although some RBSs situation-action heuristics evident in the thinking-
do perform searches, most rely mainly on a representa- aloud problem-solving protocols of experts. Second,
tion of the problem and chunks of the solution to sim- researchers have been able to develop learning pro-
plify a task. Search is more of a last resort for these cedures capable of inferring rules from experience.
problem solvers. In fact, most RBSs perform little or no RBSs can thus often accept and assimilate newly
search. learned rules merely by incorporating them into the
knowledge base.
Programming Specialized RBS architectures have evolved to ad-
We have already stated that RBS technology requires a dress different target applications. Each specialty seems
rule writer to consider and understand the organization to benefit from a slightly different emphasis. Today,
and operation of the target system. Rule writing can in special formalisms and supporting systems have been
fact be described as a special kind of programming. Like developed in such areas as
conventional programming, effective rule programming
requires mental modeling of state changes, syntactic 1. rule-based programming,
and semantic checking of rule conditions and execution 2. rule-based signal understanding,
effects, and heuristic methods for validating and verify- 3. rule-based cognitive simulation,
ing a proposed system. In contrast to conventional pro- 4. teachable and learnable RBSs,
gramming, however, rule-based programming requires 5. systems for learning rules, and
an author to think more analytically than procedurally,
Most programmers have some difficulty with this for a 6. systems for building commercjal rule-based expert
few weeks. Instead of first appreciating the relevant systems.
goals and heuristic methods and then implementing a By now, the key ideas of RBSs have been incorporated
corresponding customized problem-solving program, in such other areas of computing as
rule-based programmers must first understand the gen-
eral method of rule-based problem solving and then 1. rule-based subsystems for communications archi-
describe a problem and its related heuristic methods in tectures;
a form consistent with the available knowledge base 2. rule formalisms for representing military doctrine,
and inference engine. This is a different skill entirely. standard policies, and historical precedents;
The rule programmer must formulate the heuristics 3. rule-based controlled deduction;

September 1985 Volume 28 Number 9 Communications of the ACM 927

Special Section

Fact
Rule memory memory

fnactive Inactive
Activation/ rules facts
tivation

Active Active
rules facts

outputs 4g-

Data Updates

Rule
antecedent

Unfinished +

r--l< actions Data-flow

network
for
partially evaluateci
Delete rule activations
completed
activations -

Matching
I I Selected
action <rule, data>
Incomplete pairs
procedures

Candidate
Control <rule, data>

c procedun
Agenda 4
activations

Preferences
and
priorities
t

FIGURE5. A RepresentativeSophisticatedRBS
II Metarules

928 Communicati~ms of the ACM September 1985 Volume 28 Number 9

 Special Section

to combine intermediate results dynamically limits the

flGURE5. A RepresentativeSophisticatedRBS range of applications.
Early AI problem-solving languages like PLANNER,
The basic RBS is just a knowledge base and an inference which was developed at MIT, provided a way to repre-
engine. The trend, however, is toward more complex and sent rules within the context of programmable theorem
sophisticated systems like the one pictured here. This trend
is motivated by two goals: greater language clarity and better
provers. Workers at Carnegie-Mellon were the first to
run-time parfom7ance. build RBSs with thousands of rules and to develop effi-
Knowledge clarity has to do with expressibility and intelligi- cient compilers and translators. One such RBS, known
bility. Experts must be able to convey their knowledge to an as XCON, became the first expert system to earn a
RBS as thoroughly and as efficiently as possible, and the multimillion dollar profit. It was used to eliminate er-
RBS in turn must be able to convey its knowledge and rors in Digital Equipment Corporation VAX orders. The
related reasoning to humans. Knowledge clarity also facili- general rule-based programming system known as OPS,
tates the modifrcatii and extension of knowledge bases. which was used for XCON, has since been used for
Many of the features evident in the diagram, such as the several other RBS applications.
multidimensional working memory (which distinguishes such Workers at Stanford developed the MYCIN family of
dimensions as space, time, or level of abstraction) and the
separation of rules from metarules and control procedures,
RBSs. MYCIN was the first RBS the expertise of which
improve knowledge clarity. was acknowledged as such by experts. It was able to
The performance goal also motivates many of the embel- perform expert-level subjective reasoning with uncer-
lishments of the sophisticated system. A rule compiler con- tain data and knowledge and to explain its reasoning in
verts the triggering data conditions into a data&w network English. A similar system called PROSPECTOR, which
that optimizes the computing required to identify executable was developed by SRI to automate knowledge of min-
rules. The sophisticated system exploits several additional eral deposits, is credited with producing multimillion
mechanisms that help to determine which rule should ba dollar benefits for at least one mine operation. Subse-
executed next. The prioritized list of rules awaiting execution quent work at Stanford on TEIRESIAS and MRS empha-
constitutes the agenda. Higher level rules known as meta-
sized metarules for expressing explicit knowledge
rules express preferences that can influence the priority of
specific candiiates on the agenda. The schedulsr examines
about control.
the agenda of waiting rules and considers the applicability of Systems at Stanford and CMU that were originally
any specialized control procedures the system includes. It developed to reason about signal data have evolved to
then selects either a new procedure, a procedure continua- handle large macrorules known as specialists, knowledge
tion, or the action of a high-priority rule for execution. sources,or pattern-directed modules. These systems,
among which are HEARSAY-II, HASP, AGE, and BB-1,
often pack a great deal of knowledge into a single mod-
ule. Each module has a condition and an action, and as
4. pattern-directed modules, or macrorules, for dis- overall systems, they behave like other RBSs we have
tributed architectures and systems of cooperating considered. They differ from more typical RBSs in us-
experts; ing local memory in their computations. In this regard,
5. metarules for heuristic adaptive control of resource- they have much in common with object-oriented archi-
limited systems; tectures like Smalltalk and Ada@packages.
6. rules as a basis for enforcing constraints. PROLOG was the first general-purpose logic-based
programming language. PROLOG is essentially an RBS
THE EVOLUTION OF RBS TECHNOLOGY that uses stored facts and rules to deduce solutions to
RBS technology incorporates many ideas from diverse goal patterns. It was designed for theorem proving, but
sources. A brief and highly simplified recounting of this has proved attractive for a wider range of AI tasks.
development follows, focusing on the principal develop- The RITA and ROSIE systems developed by RAND
ments in computer science that have most advanced advance the use of RBS methods for conventional pro-
the RBS field. gramming. These systems blend rule-based program
The starting point was decision tables and compilers. representation with flexible I/O. They are thus very
This technology, which emerged about 20 years ago, attractive for designers of automated intelligent assis-
provided a representation of decision logic for transac- tants for computer-based communication tasks.
tion processing and report generation. Decision table The M.le programming system developed at
entries define condition-action rules that execute se- Teknowledge incorporates techniques for tolerating un-
quentially on the current input data. The context ef- certainty and combining evidence in a general-purpose
fects are immediate because there are no working data. rule-based programming system that operates on an
The only knowledge-base entries are the rules, which IBM personal computer. M.l marries the rule-based
must represent simple Boolean conditions. The short- programming capabilities of PROLOG, RITA, and
comings of decision tables were, in retrospect, consider- ROSIE to the evidence-combining capabilities of
able: Large rule tables are quite complex, the rigid or- MYCIN.
der of rule evaluation often proves unsatisfactory, and Ada is a trademark of the U.S. Department of Defense.
the inability to describe complex symbolic patterns or Ml is a trademark of Teknowledge.

September 1985 Volume 28 Number 9 Communications of the ACM 929

Special Section

RBS technology is the result of efforts to apply general con- perform all computations. Subsequently, theorem provers
cepts from psychology and computing theory to the simula- were developed for automating deduction. The dual goals of
tion of expertise. This figure, which is in the form of a nauti- performing human problem-solving tasks and avoiding some
lus, shows how RBS technology, psychology, computing of the gross inefficiencies of general-purpose deduction led
theory, and various application areas are interconnected. to condition-action production-rule systems. These often
Each new development in one area is shown to depend on emphasized low-level and detailed activities, though, and so
developments in related areas, as well as on generations of knowledge rules were developed that could embody chunks
previous developments. Different pathways through the spi- of expert know-how. With the initial success of knowledge
ral illustrate (different historical perspectives on technological rules for expert systems, many people began developing
history. Tracing the spiral clockwise recapitulates the succes- general-purpose rule-based programming systems. A single
sive cycles of concurrent activity in the four sectors. Travers- rule in these systems could often combine analytical knowl-
ing a radial spoke outward from the center recapitulates the edge about a problem domain with control knowledge about
successive c:hanges in concept and zeitgeist within a disci- ways to achieve problem-solving efficiency. These two forms
pline. were subsequently distinguished, with control blocks ex-
At the outset, Markov rules provided a simple technique pressing imperative knowledge in procedural form, and meta-
for defining stochastic processes with probabilistic rules that rules representing other forms of knowledge about knowl-
mapped any c&rent state into its possible successors. Math- edge in a rule-based format.
ematical theorist Emil Post showed how machines that fol- The cognate areas have both supported and adapted to
lowed simplt? string-matching condition-action rules could developments in FIBS technology. Computing theory has

FIGURE6. The Evolution of RBS Concepts

930 Communications of the ACM September 1985 Volume 28 Number 9

 Special Section

High-level.problem-solving.approach:
In order to diagnose and repair a car, follow this procedure:
C is a car.
Display the following:
Welcome to the Car Charging Diagnosis and Repair Adviser.
Find out about a car called C.
Determine the initial symptoms of the car C.
Determine the cause of the problem with the car C.
Determine the recommendations for fixing the problem
with the car C.
Show the recommendations to fix the problem with the car C.

S.1 provides a procedural syntax for expressing imperative sophisticated RBS enhances intelligibility and produces im-
knowledge. By distinguishing control blocks and rules, the proved explanations of its lines of reasoning.

FIGURE7. A ControlBlockExpressingProceduralKnow-Howin a SophisticatedRBS

The S.le expert-system building tool, also developed as defined by the DOD-sponsored Strategic Computing
at Teknowledge, advanced previous RBS technology by Initiative, are
differentiating representations for analytic and impera-
1. increasing the size of practical rule bases to 10K or
tive knowledge. S.1 employs rules for analytic knowl-
more,
edge and procedural control blocks for imperative
2. increasing the speed by two orders of magnitude or
knowledge. It provides a built-in back-chaining control
more,
mechanism with points of escape for user-supplied con-
3. broadening the set of inference techniques used by
trol blocks. By separating these two forms of knowl-
interpreters,
edge, users can create more intelligible knowledge
4. improving the methods for reasoning with uncer-
bases that are capable of providing much improved au-
tainty,
tomated explanations. Figure 7 shows a control block
5. simplifying the requirements for creating and ex-
that defines an approach to organizing rule-based rea-
tending knowledge bases, and
soning for diagnosing automotive problems. It would
6. exploiting parallel computing in RBS execution.
employ the automotive diagnostic rules shown in
Figure 2.
ASSESSMENT: THE REPUTATION
VERSUS THE REALITY
IMPLEMENTATION AND AVAILABILITY Because RBSs have played an important role in demon-
Table II is a list of supported tools that are currently strating the importance and practicality of knowledge
available for RBSs. Current goals in RBS-related R&D, systems, they have received much attention. It should
S.1is a trademark of Teknowledge. be pointed out that RBSs are not a panacea, either for
DP problems or for AI problems. They do, however,
represent a new technology of broad and important ap-
provided a foundation in automatons, grammars, theorem plicability and will undoubtedly play an increasingly
proving, relational algebra, applicative programming styles, important role in these fields in years to come.
executable specifications, and distributed controt. Psychol- Today, rule-based components are becoming stan-
ogy has contributed some extremely general and simple dard in advanced applications. DEC and NCR have in-
views of intelligent functions (e.g., Markov models, the Ele- corporated them in their XCON and OCEAN order-
mentary Perceiver and Memorizer (EPAM), and the General
entry and configuration systems. General Motors has
Problem Solver (GPS)) to successively more knowledge-
jntensive and elaborated views of cognition. Later conceptual undertaken several rule-based expert systems for man-
generations distinguish what is known from how it can be ufacturing and service functions. Numerous aggressive
applied, taught, or made more efficient. The primary focus of development programs under way throughout the
applications has shifted from general-purpose simulation, world, including the Fifth Generation program in Japan,
string processing, and automated deduction, to more press- the Alvey program in England, and the Strategic Com-
ing, higher value, and knowledge-intensive concerns. Suc- puting and MCC programs in the United States, all aim
cessively, these tasks have emphasized the need for heuris- to make significant improvements in the performance
tic solutions, specialized expert systems for problem solving, and generality of this technology over the next five to
autonomous intelligent agents, knowledge systems for stor- ten years.
ing and distributing large quantities of institutional knowledge
We conclude by listing the key strengths and weak-
in electronic form, and heuristic systems for adaptive control
and other management tasks. nesses of RBS technology, as well as some long- and
short-term objectives. First the strengths:

September 1985 Volume 28 Number 9 Communications of the ACM 931

 Special Section

TABLE II. A Representative Set of RBS Software Products

bduct Host Vendor
Tools for commercial M.l IBM PC Teknowledge
development OPS VAX Digital
s.1 VAX, Symbolics Teknowledge

Tools for research and ART Symbol& Inference

experimentation ROSIE VAX, Xerox RAND

1. RBSs can represent problem-solving know-how in a 3. improved architectures for using metaknowledge
manner suitable for application by computers; efficiently,
2. they modularize chunks of knowledge: 4. automatic translation of diverse forms of knowledge
3. they support incremental development; into rule form, and
4. they make decision making more intelligible and 5. optimizing compilers for performing global data-
explainable; flow optimizations and exploiting multiprocessor
5. they provide a fertile framework for conceptualiz- opportunities.
ing computation in general:
6. they open new opportunities by providing a non- Acknowledgments. I am grateful to the many col-
von Neumann schema that can exploit parallelism leagues who have shared their experiences and percep-
in computer systems; tions regarding RBSs with me. While there are too
7. specialized RBS architectures have emerged that many people to name, most are at Carnegie-Mellon,
constrain and simplify application methods; Stanford, RAND, and Teknowledge. I am also grateful
8. recent advances in RBS technology distinguish im- for the generous support from Teknowledge that has
perative and analytic know-how evenas they inte- made the contribution of this paper possible.
grate them to produce more effective, cogent, and
maintainable knowledge bases; FURTHER READINGS
9. rule-based reasoning can provide a conceptual basis The suggested readings span a range from introductory to state of the
for the analytic formulation of imperative know- art. The article by Duda and Gasching is a readable. simple introduction.
Erman et al. describes the motivations behind S.1 and the techniques
how. that were used to distinguish analytical rules from imperative prescrip
tions. Hayes-Roth et al. surveys a variety of issues in representing and
RBSs are still without several features that would implementing know-how; it covers but is not limited to RBSs. Shortlifle
help to make them more suitable as a general comput- describes one of the seminal projects in RBS technology. Waterman and
Hayes-Roth surveys the RBS field and the closely related but more
ing approach; specifically, they lack general pattern-directed inference systems.
1. a precise analytic foundation for deciding which
problems are solvable, Duda, R.O., and Gaschnig, J.C. Knowledge-based expert systems coming
of age. BYTE 6, 9 (Sept. 19811, 238-278.
2. a suitab:ie verification methodology or a technique Erman. LX., Scott, A.C.. and London, P.E. Separating and integrating
for testing the consistency and completeness of a control in a rule-based tool. In Proceedings of the IEEE Workshop on
rule set, Principles of Knowledge-Based Systems (Denver. Cola.. Dec.). IEEE,
1984.37-43.
3. a theory of knowledge organization that would fa- Hayes-Roth, F.. Waterman, D.A., and Lenat. D.B. Building Expert Systems.
cilitate scaling up without loss of intelligibility or Addison-Wesley, Reading, Mass., 1983.
performance, Shortliffe. E.H. Computer Based Medical Consultations: MYCIN. Elsevier
North Holland. New York. 1976.
4. high-grade rule compilers and specialized hardware Waterman. D.A., and Hayes-Roth, F. Pattern-Direcfed Inference Systems.
accelerators, and Academic Press. New York. 1978.
5. methods for integrating easily and seamlessly into
conventional DP systems. CR Categories and Subject Descriptors: 1.21 [Artificial Intelligence]:
Applications and Expert Systems; 1.2.3 (Artificial Intelligence]: Deduc-
The near-term research objectives include tion and Theorem Proving-deduction (e.g., nafural. rule-based); 12.4 [Ar-
tificial Intelligence]: Knowledge Representation Formalisms and Metlz-
1. integration with conventional software systems, ods: 12.5 [Artificial Intelligence]: Programming Languages and Software
Additional Key Words and Phrases: rule-based systems
2. modularization and reuse of components, and
3. shareab.tlity of knowledge bases among several re-
Author’s Present Address: Frederick Hayes-Roth, Teknowledge. Inc.. 525
lated applications. University Avenue. Palo Alto. CA 94301.
Beyond these near-term objectives, some long-term
goals are Permission to copy without fee all or part of this material is granted
provided that the copies are not made or distributed for direct commer-
I. improved hardware for storage and execution tasks, cial advantage, the ACM copyright notice and the title of the publication
and its date appear. and notice is given that copying is by permission of
2. improved standard software and algorithms for the Association for Computing Machinery. To copy otherwise, or to
common functions, republish. requires a fee and/or specific permission.

932 Communications of the ACM September 1985 Volume 28 Number 9