Professional Documents
Culture Documents
Paper Pitch Problem
Paper Pitch Problem
Paper Pitch Problem
Articles
PITCH EXPERT
A Problem-Solving System for Kraft Mills
■ PITCH EXPERT was developed to make expertise PITCH EXPERT is the largest deployed knowl-
available to mill-site engineers to solve pitch edge-based system anywhere in the pulp and
problems in kraft pulp mills. These problems paper industry. The system can now be used
have been estimated to cause losses to the Cana- directly by modem by mill process engineers.
dian pulp and paper industry in excess of $80
Although previous publications (Kowalski
million each year. The design of the system took
and Gauvin 1992a, 1992b; Kowalski and
into account not only the complexity of the pro-
cess interactions and the need for accuracy and
Lebensold 1989) have reported on the promise
completeness of recommendations but also the and potential of PITCH EXPERT , the past 18
ongoing need for training mill personnel and the months have seen this promise fulfilled, and
requirement that the system be maintainable and the proof of its positive impact, obtained from
expandable without the constant involvement of this initial 2-year period of industrial use, con-
the developers. PITCH EXPERT is now accessible by stitutes a main area of focus for this article. In
modem, and the savings achieved through use of addition, a retrospective analysis of the under-
the system covered the development costs within lying reasons for success is provided in more
six months of release. depth than was previously possible. Many fac-
tors contributing to the success of the project
P
ITCH EXPERT is a large knowledge-based could be assessed only after the system had
system that analyzes and diagnoses been used for a significant length of time
problems in kraft pulp mills that are (more than a year). This article highlights these
related to pitch deposition and pitch dirt. By factors with (1) a quantification of actual sav-
making scientific expertise directly available ings, rather than mere projections of expected
to mill engineers, the system augments their results; (2) a determination of which parts of
problem-solving capability and relieves the the system design proved to be most useful or
human expert of much of the burden of rou- even crucial to its success and why; (3) an anal-
tine problem solving and training. The devel- ysis of how useful and beneficial the aspects of
opment and delivery of this system also the design related to system maintenance and
served as a case study to determine the useful- update were over an extended period of time;
ness and relevance of knowledge-based tech- and (4) an examination of the magnitude of
nology in solving real problems in the pulp the benefits that were derived from the addi-
and paper industry. tion of extra capabilities to the system beyond
PITCH EXPERT is now serving as a significant those that normally form part of a knowledge
productivity aid to industry. At the time of base (that is, conflicting information identifi-
writing, there were already 36 mills, repre- cation, compact explanation capabilities,
senting over 20 companies, registered as extensive metaknowledge).
active users. Current estimates (see Perfor- The successful development and delivery of
mance) are that these mills will save approxi- PITCH EXPERT were largely the result of certain
mately $22 million a year by using PITCH important design decisions. In particular, the
EXPERT. Its performance has been documented structures and mechanisms used to build the
in an initial 12-month study as well as by system were specialized and customized to fit
records of more recent use. To our knowledge, the specific needs of a system operating in the
… kraft mill real-life pulp-mill environment. Moreover, include the plugging of screens and cleaners,
the system was designed from the beginning deposits on the pulp machine, sticking prob-
pitch problems to be maintained on an ongoing basis with- lems on press rolls, and excessive resin.
are estimated out the involvement of senior knowledge
Economic Significance
engineers.
to cost, on It is difficult to place an exact figure on the
average, Kraft Pulp Mills cost of pitch problems, even for a given mill.
However, kraft mill pitch problems are esti-
several In the manufacture of paper, wood is first mated to cost, on average, several million
million pulped to separate its fibers. One of the pre- dollars each year for each mill in North
dollars each dominant pulp processes is done in a kraft America; in Canada alone, this figure trans-
pulp mill and consists of cooking wood chips lates to $80 million a year.
year for each at elevated temperature and pressure in the Components of the cost include sale of off-
mill in North presence of certain chemicals (alkali and grade pulp contaminated with pitch dirt, pre-
sulfide), washing the resultant brown pulp, mature replacement of machine clothing,
America … bleaching to make the pulp white, and drying time lost for cleanups, and the cost of addi-
the pulp for shipment to a paper mill. tives to control the problem. The additives
can include detergents and solvents for clean-
Pitch ing surfaces, pitch dispersants for stabilizing
Pitch, or wood resin, is the material in wood the resin in suspension, and talc for detacki-
that is insoluble in water but soluble in fying the deposits. Even if a mill is successful-
organic solvents. It usually makes up 14 per- ly controlling pitch with the use of these
cent of the weight of wood after the bark is additives, substantial savings can often be
removed and is often a sticky material. With achieved if the addition rate can be reduced.
this gluelike material passing continually
through a kraft mill, it is not surprising that Technology of Pitch Control
under some circumstances, a certain amount Many factors are important in determining
deposits on the surfaces of the process equip- whether a kraft pulp mill has pitch problems.
ment. These factors include species and storage time
In addition to the pitch, other depositable of the wood being processed, degree of bark
materials can be present, such as hard-water removal, purity and concentration of the
soaps and defoamer components. In addi- cooking liquor, thoroughness of pulp wash-
tion, other materials have a tendency to ing, use of foam-control agents, process pH,
become entrained in deposits and increase water hardness, temperature of the process,
their rate of growth. These materials include and the use of additives for pitch control.
calcium carbonate from poor white-liquor In solving pitch problems, chemical analy-
clarification, carbon particles from fly ash or sis of pitch deposits is often an important
green-liquor dregs, bark particles, and sand key. Numerous analytic chemical methods
introduced with wood chips. Although they can determine the composition of the
contain many materials besides wood resin, deposits. In a troubleshooting situation, this
the deposits are still usually referred to as analysis leads to a better knowledge of what
pitch deposits. is depositing and usually points to a course of
action to solve the problem. Mill personnel
Kraft Mill Pitch Problems can also perform a number of diagnostic tests
Kraft mill pitch problems can take a number on the pulp suspension to obtain useful
of forms. In the most frequently encountered information for determining how to solve a
problem, pitch is partially liberated from the pitch problem at a mill.
wood during pulping and tends to codeposit
with other materials on the surfaces of the Why Knowledge-Based Technology
process equipment. These deposits grow in Was Required
thickness until they reach a size at which All this information must be considered
they break away from their surfaces of attach- when attempting to diagnose a pitch problem
ment. When they break away, the chunks of in a kraft pulp mill. Furthermore, the infor-
deposit are carried with the pulp and are bro- mation and the conclusions drawn from it
ken up by the pulp agitators and pump carry various degrees of certainty, reliability,
impellers. The result is small dirt particles in and subjectivity. Missing, incomplete, and
the final product that can result in the sale of even inaccurate data are also a fact of life in
the pulp at a discounted price or even the pulp mills, as is the use of ambiguous and
loss of a customer. Other pitch problems synonymous terminology. The expert must
82 AI MAGAZINE
Articles
consider all these factors when determining dent that such a measure will have a positive … one reality
further questions and making recommenda- effect on mill production. This confidence
tions. A conventional programming language can be gained only if this same mill manager of life in a
would clearly be incapable of describing this is able to obtain a logical justification for the kraft mill is
situation adequately. action from the system. Furthermore, our
Furthermore, the questioning strategy that experience has shown that such a justification that over the
one must follow in diagnosing pitch prob- must be short and concise, so that the user years, there is
lems is dynamic in nature. One must consider can grasp the essential reasoning without get-
all the information to date before determin- ting lost in a multitude of rules and facts.
a constant
ing which question is the most appropriate Second, PITCH EXPERT was to be used as a influx of new
one to be asked next. This situation is espe- training tool. In this respect, allowing the engineers and
cially true because a given piece of informa- users to ask the system questions—such as
tion (such as species of wood, seasoning time, why is this question important, what do you technical staff
pH values) might be relevant to several possi- mean, how do I find the answers to this ques- members.
ble causes of pitch deposition (Kowalski and tion, and why do you make this recommen-
Gauvin 1992a, 1992b). dation or conclusion—is extremely impor- Thus, there is
The mechanism for handling the question- tant. Also, offering this same user the ability a constant
ing logic of the system had to be flexible to specify the level of detail with which the
enough to allow for the easy addition of more answer is displayed further enhanced this use
need for
sophisticated features at a future date, with- of the system. training.
out the need to redesign a part of the system.
Such features could include an answer-retrac- The Need for Training
tion capability (Reiter and de Kleer 1987) as of Mill Personnel
well as a mixed-initiative mode of use that The experience of the domain exper t,
would allow the user to enter information Lawrence Allen, in solving kraft mill pitch
immediately rather than wait for the appro- problems now spans over 20 years. Although
priate question (Kowalski and Gauvin 1992a; there are still some grey areas in our under-
Kowalski and Lebensold 1989). standing of the various interacting phenome-
PITCH EXPERT was designed to meet all these na of pitch deposition and dirt formation,
functional requirements and satisfy the previ- experience suggests that we have sufficient
ously mentioned constraints. However, it was practical knowledge to provide expertise for
equally important for the system to be easily solving most pitch problems in kraft mills.
maintainable and updatable so that as new Because of this existing knowledge and the
methods of pitch control are discovered, the complexity of pitch problems, a considerable
expertise could quickly and efficiently be amount of consulting time has been spent
incorporated into the system. In this way, helping mill personnel with pitch problems.
state-of-the-art expertise could always be At the mills, over time, certain personnel
accessible to mills across the country. Similar- learn through reading, consulting sessions,
ly, PITCH EXPERT could be kept up to date with and experience about how to avoid pitch
regard to changes in kraft bleaching technolo- problems. Nevertheless, one reality of life in a
gy. This requirement was carefully considered kraft mill is that over the years, there is a con-
from the initial design to the final imple- stant influx of new engineers and technical
mented version, and in retrospect, the high staff members. Thus, there is a constant need
priority accorded this issue largely contribut- for training. No textbooks currently deal with
ed to the system’s success (Kowalski and this subject.
Muise 1990).
Finally, one additional requirement necessi-
tated building the system as a knowledge
Implementation
base. PITCH EXPERT had to be able to explain PITCH EXPERT was implemented using the ART
itself clearly and concisely in relation to every (automated reasoning tool) expert system
aspect of user interaction, including questions development tool on top of the Lisp program-
it might pose and conclusions and recom- ming language. The hardware platform was a
mendations it might provide. This require- SPARCSTATION 1 with 32 megabytes (MB) of ran-
ment was important for two reasons. dom-access memory and 500 MB of hard-disk
First, before a mill manager will authorize storage. The system contains approximately
the implementation of a recommendation 1200 rules and 3000 schemata. In addition,
that might result in a significant expenditure about 200 functions and demons (procedures
(such as adding or increasing the feed rate of attached to knowledge structures) are present
a chemical additive), he or she must be confi- in the system.
FALL 1993 83
Articles
Evidential Reasoner
Ntk-module Library
O NTK-network F exp
F create-conclusions
R NTK-processor F i/o
F std F create-recommendations
S summary.art
Domain Knowledge
84 AI MAGAZINE
Articles
FALL 1993 85
Articles
KEY
ntk-fact
comes from/
leads to relation
D deamon NTKF-12601
acid number of
assertion of one of tall oil is ?value
possible NTKIF's
ASSERTS ONE
OF FACTS D comes-from
BELOW
operation of the NTK network. In the two years that have passed since PITCH
This separation of questioning logic and EXPERT was delivered, the NTK network has
control into a distinct module (Hughes 1987) undergone almost continuous modification,
and the use of a network of object-type struc- aimed at improving the questioning strategies
tures (as opposed to the implementation of of the system. These modifications have been
the logic completely in the form of rules) has performed exclusively by the Paprican mainte-
greatly simplified the task of maintenance nance team (as opposed to the senior knowl-
and development (Kowalski and Lebensold edge engineers), always using the NTK dia-
1989). As new questions were acquired from grams and method; this maintenance
the expert, they were added to the system approach has proven itself to be straightfor-
with relatively little work; all we wer e
ward and simple. Clearly, it is one of the
required to learn was the appropriate position
aspects of system design that is responsible for
of the question in the NTK network, the rele-
the continuous and smooth improvement in
vant facts that could be derived from answers
the system’s performance since its delivery.
to the question, and the relevant connections
The nodes of the NTK network also offer
between these facts and other facts already
existing in the network. an ideal centralized repository for informa-
Questioning paths or strategies can also be tion on the status of each question (that is,
changed easily because these changes need to has it been asked, can it be asked, if asked
be reflected in only one part of the system what was the answer, and so on). Without
(the NTK network). This easy mode of modifi- these structures, it would be necessary to
cation was especially important for PITCH have a set of complex and difficult-to-manage
EXPERT because many questions and the rele- rules to determine when and if it was neces-
vant information derived from their answers sary and meaningful to ask a given question.
play an important role in more than one Furthermore, these NTK nodes provide an
cause of pitch deposition. As a result, the easy attachment point for various kinds of
answers to questions related to one cause of deep knowledge and even metaknowledge,
pitch problems can dynamically affect the for example, online help facilities (what do
questioning path for a second cause. you mean, why are you asking, how do I find
86 AI MAGAZINE
Articles
Symptoms
incomplete cooking Chemical analysis Pulp mill model
excessive use of defoamer metal soaps locations
presence of contaminants resin substances
use of resinous species dispersed resin equipment
improper wood seasoning resin in bark processes
poor debarking observables
defoamer
poor brownstock washing recommendations
calcium carbonate
poor white liquor clarification
contaminants problems
improper use of additives
carbon particles tests
metal soap deposition
inappropriate use of defoamer
the answer, what are acceptable answers, and domain knowledge relating to pitch prob-
so on). These help facilities have been found lems. It consists of two major components:
to add greatly to the system’s ease of use, thus the mill model and the diagnostic module.
gaining the confidence of the pulp-mill per- Mill Model As shown in figure 6, the mill
sonnel. They have also made it possible to use model consists of five semantic networks, each
PITCH EXPERT as a training tool, which is impor- of which describes a typical kraft pulp mill
tant because of high turnover in the mills. from a particular perspective (locations, sub-
This training capability is seen by the users stances, equipment, processes, and observ-
(that is, mill personnel) as a feature of great ables). These five networks are connected, as
importance and is one of the reasons for its appropriate, at various points. Together they
enthusiastic acceptance. Our analysis of the provide a model of the pulp mill with both a
use of the system in its first two years shows a lateral and a hierarchical frame of reference
common trend among initial users to make that is critical to the proper and efficient func-
heavy use of this feature. In fact, many first- tioning of the rules in the diagnostic module.
time users have run a complete session of well In addition, three semantic networks relating
over 100 questions asking why for each one. to problems, tests, and recommendations are
Such an exercise has allowed them to gain a used in the overall reasoning process.
deeper understanding of the true complexities The eight networks combined form a mod-
of pitch-deposition problems. el consisting of over 2000 nodes (implement-
The NTK network also allows PITCH EXPERT ed as schemata) and over a dozen relations,
great flexibility in handling incomplete infor- most of them customized and complex. Fig-
mation. When a question is answered with ure 7 shows a small portion of one of these
unknown, alternate questioning paths are networks, describing the locations in a pulp
already explicitly laid out in the NTK network mill. Although not a model-based system as
and can quickly and efficiently be activated. formally defined (Winston and Shell 1990; de
Other potential features that can be imple- Kleer and Williams 1989), the mill model pro-
mented easily and efficiently using this vides some of the same advantages (de Kleer
approach include (1) multilingual operation 1991) by supporting and supplementing the
and more advanced help and explanation diagnostic heuristics of the system.
capabilities; (2) answer-retraction capability, Although the use of such a structure allows
where the truth maintenance issue is not the option of customizing the models to
affected, but the ability of the system to read- match any particular mill exactly, our experi-
just the questioning strategy automatically ence to date has shown the need for such a fit
and dynamically would be of great help; and to be less than urgent. The generic model of
(3) mixed-initiative mode capability. The the mill has served well in approximating the
installation of such a capability would be sim- configuration of the actual mills that have
plified greatly by the already-existing model of used the system. It remains to be seen if the
the relationship of each question-answer set to customization option becomes more impor-
all its immediately related facts and questions. tant as use of the system grows, and kraft pro-
cess operations change.
Domain Module Diagnostic Module The diagnostic module
The domain module contains the actual implements the diagnostic reasoning strate-
FALL 1993 87
Articles
LOCATION IS-A
PART-OF
EQUIPMENT LOCATED-IN
NEXT-DOWNSTREAM
UNBLEACHED
LOCATION
BROWNSTOCK BROWNSTOCK
COOKING WASHING SCREEN ROOM
AREA AREA AREA
DIGESTER
BROWNSTOCK
SCREENS
KNOTTERS
BLOWTANK
FIRST UNBLEACHED
BROWNSTOCK CLEANER
WASHER
BROWNSTOCK
TANK
MIDDLE
BROWNSTOCK
WASHER BROWNSTOCK
WASHED DECKER
STOCKCHEST
LIQUOR
PUMP BROWNSTOCK
WASHERS
LAST
BROWNSTOCK
WASHER
gies related to pitch problems. As shown in egy (and is comparable to message passing for
figure 6, it consists of 19 components, each encapsulated objects).
of which addresses one particular aspect of Third, inferred facts are created from the
pitch diagnosis, either symptoms of pitch diagnostic rules when the conditions of the
problems or chemical analyses performed to rules are satisfied by NTK facts and previously
diagnose pitch problems. Each component inferred facts; they serve, in turn, to satisfy
uses five types of structure to produce its other diagnostic rules (creating the inference
diagnoses: chains of the system).
First, NTK facts, which were described in Fourth, evidence structures are created by the
the preceding section, are mentioned here diagnostic rules. They represent conclusions
only because they serve to satisfy the condi- about the likelihood of a particular cause of
tions of diagnostic rules. pitch deposition based on the certainty and
Second, metaasserts are commands that reliability of one source of information. Natu-
send a message, when appropriate, to the rally, for any given cause of pitch problems,
NTK module, instructing it to track down a there might be several sources of informa-
particular piece of information that is needed tion, each of which produces a separate piece
for the diagnosis. Note that the sending of of evidence. When all the available informa-
this message is the only involvement of the tion has produced all the relevant pieces of
diagnostic module with the questioning strat- evidence, the evidential reasoner (see next
88 AI MAGAZINE
Articles
EVIDENCE-OF
EXCESSIVE DEFOAMER USAGE
NTK-FACT SOURCE : WOOD SPECIES
WEIGHT : 4
WOOD SPECIES CERTAINTY : LIKELY
IS ASPEN
INFERRED FACT
DIAGNOSTIC SERIOUS FOAMING DIAGNOSTIC
RULE PROBLEMS ARE LIKELY RULE
INFERRED FACT
INFERRED FACT
INCOMPLETE
SEASONING IS EXCESSIVE DEFOAMER
ALMOST-CERTAIN FEED RATE IS LIKELY
subsection) collates, sorts, and prioritizes reasoning, using the species of wood and
them to reach its final conclusions. information about its seasoning, is unaffected
Fifth, rules are standard, diagnostic if-then by any new lines of reasoning using other
rules that use facts (both NTK and inferred) to information. In this way, maintenance
create other inferred facts as well as pieces of requirements related to the addition of new
evidence. lines of reasoning are kept to a minimum.
Figure 8 shows a portion of the rule set per- Figure 9 diagrams a second line of reason-
taining to the problem of excessive defoamer ing from the same rule set, which uses infor-
use. This example shows how the various mation from the NTK module related to the
knowledge representation structures and feed rate and the location at which defoamer
modules of PITCH EXPERT work together in a was added, as specified by the user, to pro-
coordinated fashion. duce an inferred fact and an instance of evi-
A message is sent to the NTK module indi- dence with the appropriate parameters.
cating the need to know the species of wood
being used. If the species is aspen, and it has Evidential Reasoner Module
been inferred in another rule set that the Because of the functional requirements of the
wood is incompletely seasoned, then it will system (see Why Knowledge-Based Technolo-
be inferred that serious foaming problems are gy Was Required), it was clear early on in the
likely to occur. If so, a second rule will infer design of P I T C H E X P E R T that a standard
that it is likely that excessive amounts of approach to dealing with uncertainties, such
defoamer are being used. This same rule also as the certainty factors of MYCIN or the imple-
creates an instance of evidence specifying the mentation of the Dempster-Schafer theorem,
appropriate weight and certainty. Both of the would have been inadequate (Buchanan and
inferred facts created in this rule set can be Shortliffe 1985). Instead, a customized evi-
used to satisfy rules in other rule sets. The dence-handling strategy was developed.
instance of evidence (along with all other Figure 10 shows a typical piece of evidence,
instances of evidence) is processed at the which consists of a specification of an associ-
appropriate time by the evidential reasoner. ated cause of excessive pitch deposition, the
This example clearly shows how this line of source of the information (the specific piece
FALL 1993 89
Articles
FROM NTK
MODULE
TO EVIDENTIAL REASONER
NTK-FACT
DEFOAMER EVIDENCE-OF
USED IN
UNBLEACHED EXCESSIVE DEFOAMER USAGE
LOCATION SOURCE : FEED RATE
WEIGHT : 1
CERTAINTY : LIKELY
DIAGNOSTIC
RULE
90 AI MAGAZINE
Articles
FALL 1993 91
Articles
92 AI MAGAZINE
Articles
FALL 1993 93
Articles
printout from the database of which facts incomplete session for later retrieval.
exist as input (left-hand side of a rule) and Missing information is handled with two
which as output (right-hand side of a rule) keywords, later and unknown. Later is used
and checking this printout against the dia- when an answer is expected to be obtainable
grams. All this checking virtually assures that but is not available right now and unknown
the blueprints represent the actual state of when no better answer can ever be expected.
both the database and the code. When answers become available for questions
The success of this approach is attested to by originally answered by later, the session can
the significant improvements in accuracy and be retrieved and rerun, and these questions
completeness (as outlined in Performance) are asked again.
that have been achieved by changes made by Each answer, as it is processed, can create
Paprican personnel with minimal supervision NTKFs and, indirectly, NTKIFs. In turn, the
by the senior designers. Considering the day- NTKFs and NTKIFs activate more NTKs, or
to-day use that is still being made of this main- they can lead domain rules to fire, creating
tenance system to maintain and upgrade the inferred facts or pieces of evidence. In either
knowledge base, its importance to the success case, the chain of rules eventually leads to
As the of the project cannot be overstated. the creation of a collection of evidence to be
This maintenance mechanism has with- evaluated by the evidential reasoner.
system was stood the test of real-life use and still serves At any point in a PITCH EXPERT session, the
fine tuned today as the control and self-checking mech- conclusions and recommendations that are
anism that has allowed an ongoing dynamic supported by the evidence asserted to this
with each environment to be created in which constant point can be viewed on the user’s screen by
new mill test modification of the knowledge serves as one use of the recc keyword. These conclusions
run, the of the system’s attractions rather than a and recommendations are, of course, not
detraction. Response from the users has final and change during a session to reflect
system’s shown that the ability to have state-of-the art new information as it becomes available.
performance expertise always available is a strong source of Three kinds of help are available for ques-
motivation to use it. tions and can be accessed with keywords.
was The perception and development of PITCH What provides a reworded version of the
continuously EXPERT as an ongoing evolving system, rather question in case the original was unclear.
than a static entity, was a key factor in its suc- How gives further detail on how to obtain the
improved. cess. The provision of such a maintenance information required to answer the question.
At the time capability, in whatever form proves workable, Why expands on why it is important to ask
should be given serious consideration in the this question in the context of pitch control.
of this development of any major industrial knowl- Because a typical session can take an hour
writing, it edge base. or more, a facility has been provided (with
stands at 90- the save and restart keywords) to save the cur-
rent session at any point to be picked up
percent Use of PITCH EXPERT again at a later time. A common use of this
accurate and PITCH EXPERT resides on a SPARC workstation at feature would be to store a session in which
Paprican and is accessed by modem. The sys- later answers were given to one or more ques-
93-percent tem operates in a question-and-answer mode, tions, with a view toward retrieving and con-
complete. posing questions to the user one at a time tinuing the session when answers to these
and accepting typed answers until it has all questions became available.
the information it needs. At important deci- When PITCH EXPERT has an answer to all the
sion points during a run, PITCH EXPERT also dis- questions it judges necessary to its reasoning,
plays numbered menus of available options. it completes its reasoning process and then
Each answer is checked for acceptability as offers to display on the screen the conclu-
soon as it is entered. The criterion can take sions it has reached and the recommenda-
the form of a list of specific acceptable tions it has made. Conclusions and recom-
answers, a class specification, or a requirement mendations are shown one by one, in order
that the answer be a number. If an unaccept- of priority, each with its supporting evidence
able answer is entered, PITCH EXPERT displays a in order of reliability.
message and then asks the question again. Recommendations are displayed in a simi-
Any question can also be answered with lar manner. By typing why in response to a
one of PITCH EXPERT ’s keywords. These key- recommendation display, the user can obtain
words activate special features such as online a summary of the reasoning leading to this
help, access to intermediate conclusions and recommendation.
recommendations, and the saving of an PITCH EXPERT can produce a summary report
94 AI MAGAZINE
Articles
of the session just completed or of any earlier sons stemming for the most part from severe
completed session. This report includes all constraints on resources.
conclusions reached, recommendations The three completed evaluations show a
made, and conflicts noted. The report is dis- combined annual savings of almost $1.9 mil-
played on the screen and incorporated into lion, for the most part because of an overall
the transcript of the session. reduction in the amount of pitch-contaminat-
ed pulp produced as well as a decrease in the
quantities of additives used. This figure is
Performance especially impressive in light of the fact that it
There are two fundamental measures of represents savings for just three kraft pulp
expert system performance: the quality of the mills. With more than 40 such mills in Cana-
expert reasoning and the benefits (economic da alone, these savings almost equal the $2.24
and other) provided by use of the system. million development cost of PITCH EXPERT to
The most basic requirement of expert sys- date. The 36 mills already using PITCH EXPERT
tem performance is that it reproduce accept- can expect to save $22.4 million a year.
ably well the reasoning of the human expert.
In the case of PITCH EXPERT, its conclusions and Discussion of the Three Completed
recommendations must be accurate (that is, Evaluations
the same ones the human expert would have Given the importance that these savings rep-
come to in the same situation) as well as com- resent, it is worth examining them in greater
plete (nothing the human expert would have detail. In the interest of brevity, we focus on
concluded or recommended is left out). those mills whose evaluations of the effects of
PITCH EXPERT were complete. For reasons of
Initial Mill Trials confidentiality, the company names and loca-
To evaluate the benefits of the system, finan-
The three
tions of the particular mills cannot be
cial and otherwise, a set of initial mill trials specified. They are referred to as mills 1, 2, completed
were conducted from May 1991 to April 1992. and 3. They are, however, real and certainly evaluations
The trial runs were conducted with the coop- typical kraft pulp mills, and the sessions of
eration of 13 kraft pulp mills across Canada PITCH EXPERT that they ran involved real data
show a
(Allen and Kowalski 1992; Turney 1992). representing real pitch problems. The cost combined
During this test period, the sessions were savings are outlined here, and the results are
conducted with the aid of an intermediary at summarized in table 1.
annual
Paprican. This person was responsible for Case Study of Mill 1: Mill 1 has an annual savings of
manning the telephone and acting as the production rate of 170,000 tons of pulp. The
contact between the mills and the system. production of off-target (that is, pitch-con-
almost $1.9
When mill staff members called in, they taminated) pulp represents a loss of $40 a million …
would verbally convey the relevant informa- ton. The defoamer used in this mill costs
tion about their particular mill to the inter- $0.88 a kilogram.
mediary, who, in turn, would enter all the Before running a diagnosis with PITCH
information into the system. The correspond- EXPERT, this mill had 5 percent of the pulp it
ing output would be checked for accuracy produced contaminated with pitch. In addi-
and completeness by Allen and then would tion, it was using 3.8 kilograms of defoamer
be sent out (with corrections if necessary) to for each ton of pulp.
the mill. With the results of each session, After running a full diagnostic session with
modifications were made to the knowledge PITCH EXPERT, entering all the required infor-
base to address any errors or inconsistencies mation and implementing the resulting rec-
that were discovered. ommendations, mill 1 had reduced its off-tar-
The early sessions achieved an accuracy of get pulp to just 1 percent of total production.
60 percent and a completeness of 70 to 80 per- In addition, the rate of defoamer use was
cent. As the system was fine tuned with each reduced to 2.8 kilograms for each ton of pulp.
new mill test run, the system’s performance Therefore, the annual savings from the
was continuously improved. At the time of this reduction in off-target pulp is $272,000. The
writing, it stands at 90-percent accurate and annual savings from the reduced use of
93-percent complete. Further progress is defoamer is $150,000. Thus, the total yearly
expected. All 13 mills were to perform a com- savings is $420,000.
plete evaluation of the benefits of the system. Case Study of Mill 2: Mill 2 has an annual
As of October 1992, three had completed their production rate of 360,000 tons of pulp. The
evaluation, seven more were in progress, and production of off-target pulp represents a loss
three had suspended the study for various rea- of $40 a ton. The defoamer used in this mill
FALL 1993 95
Articles
Mill Number 1 2 3
Off target before 5% 2% N/A
Off target after 1% 0% N/A
Annual production (tons) 170,00 360,00 450,00
Annual estminated savings $420,000 $871,000 $593,600
96 AI MAGAZINE
Articles
educational purposes rather than troubleshoot- The session transcripts also serve as a
ing, including heavy use of the help facilities record of the problem-solving process. This
to learn why the system asks its questions and record can be important because pitch prob-
why it makes its recommendations as well as lems in a given mill can be intermittent, and
what-if exercises using hypothetical data. Such a successful solution procedure can be forgot-
use illustrates to the user just how complex ten by the time the problem recurs.
and difficult accurate pitch-problem diagnosis
really is. This period lasts anywhere from sever-
al weeks to several months.
Conclusions
Second, following the initial exploration of The PITCH EXPERT project is a success. The sys-
the system, a more methodical use begins with tem is now used by an ever-growing number
real-life data pertaining to the particular mill. of kraft pulp mills from many companies all
Gathering these data can often take several over Canada. Its possible use by mills in other
weeks, depending on the availability of person- countries is now being considered; some com-
nel at the mill. As more data are entered into panies outside Canada have already expressed
the system, the updated recommendations are interest.
usually produced with stronger certainty and The analysis of system performance shows
are perused in more detail, making heavy use that knowledge-based technology can indeed
of the explanation facilities. have a dramatic impact on productivity and
Third, over the long term, roughly half of costs. The savings realized by mills using
the users access the system at least once a PITCH EXPERT and following its recommenda-
month and half less often. tions clearly justify the expense of develop-
To date, PITCH EXPERT users have all been ment. Although the continued introduction
members of Paprican. As members, they are of this technology into the pulp and paper
considered to have paid for development of industry in the form of other knowledge-
the system with their industrial membership based systems will have to proceed in a
fees. Making PITCH EXPERT available for use by methodical, stepwise manner, it is clear that
nonmember companies is now being consid- the success of PITCH EXPERT has realized the
ered. There has already been strong interest first step toward this goal.
from several countries. Although the fee struc- From the technical standpoint, it is clear
ture has not yet been determined, it is clear that the success of the project was a result of
that if nonmember companies are given access the philosophy of custom fitting the system
to the system, they will be charged for its use. to the knowledge (Bouchard et al. 1991) as
well as the careful attention that was paid to
Other Benefits of PITCH EXPERT the issue of ongoing maintenance. This pro-
A number of less quantifiable but equally real cess involved the selection of a large, sophis-
benefits have also been realized in participat- ticated, and powerful shell (in this case ART)
ing mills. PITCH EXPERT provides training in followed by customization to achieve a per-
pitch-control techniques to mill engineers. fect fit with the needs prescribed by the
Many pulp mills are in less desirable remote domain knowledge.
locations and tend to be staffed by engineers Finally, PITCH EXPERT serves to highlight the
who are inexperienced or new to the industry fact that large and sophisticated expert sys-
and have a high turnover rate. The help and tems can and do provide distributed and up-
explanation facilities are particularly useful to-date expertise in a readily available and
here. During development of the system, the accessible fashion, which translates into
expert commented that when he visits mills, improved productivity and a more competi-
he spends much of his time educating mill tive industry.
personnel as well as investigating pitch prob- The major achievements of the PITCH EXPERT
lems. It is important that PITCH EXPERT be able project can be summarized as follows:
to perform both these functions. First, a large knowledge-based system is
Using PITCH EXPERT also encourages mill per- being regularly used by mill personnel to
sonnel to do more information gathering and solve real industrial problems. Few systems in
testing on a regular basis, thus improving the pulp and paper industry reach this stage
their overall problem-solving capability. The of practical usefulness. Mills already using the
sequence of questions serves to highlight system can be expected to save a total of
information that should be obtained when $22.4 million a year.
solving pitch problems, and the help facilities Second, a sophisticated question-asking
instruct users on what to look for in the mill mechanism (the NTK module) was developed
and how to perform tests and procedures. to enable PITCH EXPERT to obtain the informa-
FALL 1993 97
Articles
tion it needs without asking unnecessary or In Proceedings of the Ninth National Conference
irrelevant questions. on Artificial Intelligence, 842–848. Menlo Park,
Third, a flexible strategy was developed for Calif.: American Association for Artificial Intelli-
combining pieces of evidence of various gence.
strengths and weights to reach conclusions de Kleer, J., and Williams, B. C. 1989. Diagnosis
about the existence and importance of prob- with Behavioral Modes. In Proceedings of the
lems. Eleventh International Joint Conference on
Fourth, a powerful set of customized rela- Artificial Intelligence, 1324–1330. Menlo Park,
tions and knowledge structures was devel- Calif.: International Joint Conferences on Artificial
oped for modeling an industrial process. Intelligence.
Fifth, thanks to its maintenance-oriented
Hughes, S. 1987. Domains, Tasks, and Questions:
design and associated strategy, PITCH EXPERT is
An Approach to Knowledge Analysis. In Proceed-
being successfully maintained with minimal
ings of the Third International Expert Systems
involvement by AI specialists.
Conference, 37–44. Medford, N.J.: Learned Infor-
Acknowledgments mation Inc.
The success of PITCH EXPERT has been the result Kowalski, A., and Gauvin, D. 1992a. P I T C H
of contributions by many people in addition EXPERT —An Engineered Collection of Specialized
to the authors of this article. Christine Knowledge Structures. In Proceedings of the Ninth
Lapointe of Paprican spent many long hours IEEE Conference on Artificial Intelligence for Appli-
conducting the initial sessions with the mills cations, 20–27. Washington, D.C.: IEEE Computer
and reviewing the session log files as well as Society.
testing and debugging the system. Ana Luque Kowalski, A., and Gauvin, D. 1992b. PITCH EXPERT: A
of the Centre de Recherche Informatique de Productivity Improvement Tool for Pulp Produc-
Montreal (CRIM) contributed greatly to the tion. In Proceedings of the Eighth International
coding and testing effort, particularly for the Conference on CAD-CAM: Robotics and Factories of
NTK module. Julian Lebensold of CRIM the Future, 1854–1866. Metz, France: International
helped to get the project under way. John Society for Product Enhancement.
Opala, Jennifer Muise, and Daniel Gauvin of Kowalski, A., and Lebensold, L. 1989. A Diagnostic
CRIM contributed great effort and skill in pro- Aid to Pulp Production. In Applications of Artificial
gramming various parts of the knowledge Intelligence VII, 857–866. Bellingham, Wash.:
base. Joanne Plamondon of Paprican tirelessly International Society for Optical Engineering.
and cheerfully processed countless revisions
Kowalski, A., and Muise, J. 1990. Toward a Self-
of this article, and Virginia Bryce of Paprican
Managing Knowledge Base. In Proceedings of the
did likewise for the user manual. Finally,
Third International Symposium on Artificial Intelli-
many thanks to Renato De Mori for his most
gence, 117–123. Mexico, D.F.: Editorial Limusa.
useful advice on numerous occasions and to
Alun Preece for his helpful review of this arti- Millet, C. 1989. A Study of the Knowledge Required
cle. for Explanation in Expert Systems. In Proceedings
of the Fifth IEEE Conference on Artificial Intelli-
References gence Applications, 83–96. Washington, D.C.: IEEE
Allen, L. H., and Kowalski, A. 1992. PITCH EXPERT Computer Society.
System: A Case Study for a Knowledge-Based Sys- Reiter, R., and de Kleer, J. 1987. Foundations of
tem for Pitch Control in the Pulp and Paper Indus-
Assumption-Based Truth Maintenance Systems. In
try. In Proceedings of the Forest Sector 2000 Con-
Proceedings of the Sixth National Conference on
ference, 98–103. Ottawa, Ontario, Canada:
Industry, Science, and Technology Canada. Artificial Intelligence, 183–188. Menlo Park, Calif.:
American Association for Artificial Intelligence.
Bouchard, D. C.; Kowalski, A.; Larin Y.; and Vadas,
O. 1991. Acquiring and Structuring Expert Knowl- Suthers, D.; Woolfe, B.; and Cornell, M. 1992. Steps
edge. In Proceedings of the Tappi Engineering Con- from Explanation Planning to Model Construction
ference, 81–93. Atlanta, Ga.: Technical Association Dialogues. In Proceedings of the Tenth National
for the Pulp and Paper Industry. Conference on Artificial Intelligence, 24–30. Menlo
Buchanan, B., and Shortliffe, E. 1985. Rule-Based Park, Calif.: American Association for Artificial
Expert Systems. Menlo Park, Calif.: Addison Wesley Intelligence.
Publishing Co.
Turney, P. 1992. PITCH EXPERT: Canadian AI Success
Cawsey, A. 1991. Generating Interactive Explana-
Stories. Canadian Artificial Intelligence 30:14–17.
tions. In Proceedings of the Ninth National Confer-
ence on Artificial Intelligence, 86–91. Menlo Park, Winston, P. H., and Shell, A. R. 1990. Artificial Intel-
Calif.: American Association for Artificial Intelli- ligence at M.I.T.: Expanding Frontiers. Model-Based
gence. Reasoning: Troubleshooting. Cambridge, Mass.: The
de Kleer, J. 1991. Focusing on Probable Diagnoses. MIT Press.
98 AI MAGAZINE
Articles
Lawrence H. Allen is the director of ■ Making Sense of Gigabytes: A System for Knowledge-Based Market Analysis /
research in chemical sciences at the Tej Anand and Gary Kahn
Pulp and Paper Research Institute of ■ TPF Dump Analyzer / R. Greg Arbon, Laurie Atkinson, James Chen, and Chris A. Guida
Canada. He received his B.S. in phys-
■ Marvel: A Distributed Real-Time Monitoring and Analysis Application / U. M. Schwuttke,
ical sciences at Carlton University in
A. G. Quan, R. Angelino, C. L. Childs, J. R. Veregge, R. Y. Yeung, and M. B. Rivera
1965 and a Ph.D. in chemistry from
Clarkson University, Potsdam, New ■ Pharos—The Single European Market Adviser /
York, in 1970. Since 1972, his research has focused Ebby Adhami, Michael Thornley, and Malcolm McKenzie
on pitch control for both pulp and paper mills, ■ The Credit Assistant: The Second Leg in the Knowledge Highway for American Express /
making him one of the paper industry’s most James Dzierzanowski and Susan Lawson
sought-after consultants. He is a fellow of the
■ MOCCA: A Set of Instruments to Support Mortgage Credit Granting /
Chemical Institute of Canada and the Technical
Steve Hottiger and Dieter Wenge
Association for the Pulp and Paper Industry.
■ Cresus: An Integrated Expert System for Cash Management / Pete Shell, Gonzalo Quiroga,
Juan A. Hernandez-Rubio, Eduardo Encinas, Jose Garcia, and Javier Berbiela
Yves Larin is an associate technical
specialist at the Pulp and Paper ■ DMCM: A Knowledge-Based Cost-Estimation Tool /
Research Institute of Canada. He Norman Crowfoot, Scott Hatfield, and Mike Swank,
received his B.S. from University of ■ SlurryMinder: A Rational Oil Well Completion Design Module /
Montreal in 1979. He has spent the E. Brent Kelly, Philippe Caillot, Robert Roemer, and Thierry Simien
last three years working in the field
■ An Application of Model-Based Reasoning in Experiment Design /
of knowledge-based systems.
Andrew B. Parker and W. Scott Spangle
■ A Truly Magic Solution / Rita C. Kidd and Robert J. Carlson
Oliver M. Vadas is a senior technical
■ Adjudipro / J. P. Little and Mark Gingrich
specialist at the Pulp and Paper
Research Institute of Canada. He ■ HUB SIAASHING:A Knowledge-Based System for Severe, Temporary Airline Schedule
received a B.S. in electrical engineer- Reduction / Trish Dutton
ing and an M.S. in computer engi- ■ Arachne: Weaving the Telephone Network at NYNEX / Elissa Gilbert, Rangnath Salgame,
neering in 1963 and 1965, respec- Afshin Goodarzi, Yuling Lin, Sanjeev Sardana, and Jim Euchner
tively, from the University of
■ Knowledge-Based Code Inspection with ICICLE /
Technology of Budapest, Hungary. He joined the
Pulp and Paper Research Institute of Canada in L. Brothers, V. Sembugamoorthy, and A. Irgon
1972, where he works on evaluating emerging tech- ■ Automatic Programming for Sequence Control / Hiroyuki Mizutani, Yasuko Nakayama,
nologies for industrial applications. Satoshi Ito, Yasuo Namioka, and Takayuki Matsudaira
FALL 1993 99