Knowledge Representation Methods for Dairy Decision Support

OUR INDUSTRY TODAY
Knowledge Representation Methods for Dairy Decision
Support Systems
H. HOGEVEEN,1.2 M. A. VARNER,, D. S. BREE: D. E. DILL,5
E. N. NOORDHUIZEN-STASSEN,' and A. BRAND1
Department of Herd Health and Reproduction
Utrecht University
Yalelaan 7
3584 CL Utrecht, The Netherlands
and
Department of Animal Sciences
University of Maryland
College Park 20742
and
Computer Science Department
University of Manchester
Manchester M 1 3 9PL, England
and
CenexRand OLakes Ag Services
PO Box 64089
Mail Station 670
St. Paul, M N 55164-0089
ABSTRACT
edge of how to use declarative knowledge. For both types of knowledge,
several characteristics can be defined:
completeness, certainty, generality, and
level. Knowledge representation schemes
can be ranked according to their performance on the various knowledge characteristics.
Common schemes for knowledge
representation and their strengths and
weaknesses are described. Different
knowledge representation schemes are illustrated for mastitis and reproductive
management.
(Key words: knowledge-based systems,
knowledge representation, dairy farm
management, decision support systems)
Knowledge-based systems are currently being applied for decision support
systems for management of dairy farms.
An important feature in the development
and application of knowledge-based systems is the knowledge representation
scheme used. Although many knowledge
representation schemes are available in
artificial intelligence, the existing dairy
farm management applications only use
production rules. However, the knowledge required for dairy farm management may require other representation
schemes, depending on the type of
knowledge involved in the decisionmaking process.
Two classes of knowledge can be distinguished: declarative and procedural (or
operational) knowledge. Declarative
knowledge is concerned with facts in a
domain. Procedural knowledge is knowl-
Abbreviation key: BBN = Bayesian belief
network, CCM = conditional causal model,
KBS = knowledge-based system.
INTRODUCTION
Received April 18, 1994.
Accepted July 5 , 1994.
'Utrecht University.
*Current address: Institute of Environmental and
Agricultural Engineering (IMAG-DLO),
PO Box 43, 6700
AA Wageningen, The Netherlands.
3University of Maryland.
4University of Manchester.
SCenex/Land O'Lakes Ag Services.
1994 J Dairy Sci 77:3704-3715
Dairy farm management encompasses a
wide range of activities (5). Decisions concerning these activities are made continually, often
using imprecise methods or incomplete information, which can result in suboptimal results.
Computer-based tools can be used to support
the decision process, thus enhancing the impact and result of the decision (62). Ration
balancing and sire selection are two examples
3704
OUR INDUSTRY TODAY
3705
of areas in which computer-based tools have Declarative Knowledge
gained widespread use. A relatively new techAt the farm level, declarative knowledge
nique in decision support systems is the use of
includes
observations made on a farm or
knowledge bases. Use of these emerging apknown
to
be important for the farm enterprise.
proaches can benefit the user and the developer
by providing more robust applications that are These observations can be made by persons or
by automatic devices (sensors). Data collected
less costly to maintain.
The method by which knowledge is automatically are sometimes preprocessed berepresented in a knowledge base, i.e., the fore they are stored and ready for use. Declaraknowledge representation scheme, can vary. tive knowledge can be available either from an
Many different knowledge representation on-farm or off-farm database. Two important
schemes have been or are being developed for features of declarative knowledge can be
specific domains. Every knowledge representa- described in terms of completeness and certion method has its own strengths and weak- tainty. When decisions are being made on a
nesses with respect to the characteristics of the dairy farm, knowledge is often incomplete,
knowledge to be modeled. The knowledge uncertain, or both.
Completeness. Incompleteness refers to the
representation scheme chosen for a decision
support system has consequences for the per- proportion of observations that are missing.
formance of the system and the execution of For example, when observations are made usits task (1 1). Knowledge involved in one aspect ing automated sensors (e.g., electrical conducof dairy farm management often has charac- tivity measurements of the milk), malfunctionteristics different from knowledge in other ing equipment can lead to missing
parts. Consequently, knowledge for a decision observations.
support system for dairy farms may require
Certainty. Observations can be regarded as
more than one representation method.
certain. However, during decision making, obVarious authors (19, 22, 23, 30, 31, 55, 57) servations often have to be translated into
have described the use of knowledge-based more general terms that are used as a base to
systems (KBS) in agricultural management. draw inferences. For example, in disease diagAlthough various knowledge representation nosis, it is useful to know whether an animal
schemes exist, most descriptions are limited to has a fever, a body temperature >39"C. A
rule-based systems. No overview is available difference of only .2"C (38.9"C vs. 39.1"C) can
of the characteristics of knowledge representa- distinguish between an inference of fever or no
tion schemes in relation to their possible use in fever. A decision maker would be more certain
dairy farm management support.
of a fever from a sensor reading of 43'C than
The objectives are therefore to describe var- from a reading of 39.1'C. Thus, representation
ious knowledge representation methods with of knowledge using fuzzy boundaries is often
their strengths and weaknesses and to provide better, e.g., 39.1"C would be classed as a fever
examples of their application to dairy farm with less certainty than 43°C.
management. Emphasis is on methods for
which product development tools are available.
Procedural Knowledge
KNOWLEDGE CHARACTERISTICS
Knowledge includes facts about the problem and a wide array of problem-solving
strategies that an expert accumulates over time
(10, 59). To solve a problem, two classes of
knowledge are often necessary: 1) knowledge
about facts in the domain, declarative knowledge, and 2) knowledge of how to use this
declarative knowledge, procedural or operational knowledge (28). Both types of knowledge have their own features and are described
herein.
Procedural knowledge uses observations
and data and transforms them into information
that is expected to be useful to the user. The
procedural knowledge in conventional computer programs, such as a management information system, rearranges and combines the
various observations to make them easier to
interpret. The procedural knowledge in KBS
seeks to help solve specific problems (63).
Procedural knowledge has three characteristics:
1) generality, 2) certainty, and 3) knowledge
level.
Journal of Dairy Science Vol. 77, No. 12, 1994
3706
HOGEVEEN ET AL.
Generality. Procedural knowledge can be
general or specific. Specific knowledge can
consist, for example, of associations between a
problem situation and a solution. These associations are often developed from experience and
are sometimes described as rules of thumb. An
example of associative knowledge is the use of
antibiotic therapy. An experienced veterinarian
knows what antibiotic to prescribe in which
situation without thinking about the exact
mechanism of action of that antibiotic. Once
such heuristic knowledge is represented, it can
be utilized repeatedly without understanding
the underlying mechanisms of action.
In contrast to heuristic knowledge, generic
knowledge consists of a causal explanation for
the undesirable characteristics of that situation.
This explanation permits the assignment of
ultimate causes and the elucidation of pathways leading from those causes to the situation
characteristics (60). The knowledge used to
generate a causal explanation of a problem is
more general; Le., the knowledge can also be
used to explain the workings of a system or to
simulate the behavior of a system.
Certainty. As with declarative knowledge,
certainty is also an important feature in procedural knowledge. Reasoning under uncertainty
is common in disease diagnosis. Disease diagnoses by experts often takes a qualitative form,
including uncertainty (e.g., Staphylococcus
aureus is the most likely pathogen causing this
specific mastitis case). The decision processes
underlying these diagnoses consist of procedural knowledge with varying degrees of uncertainty. Uncertainty is common in biological
systems in which precise knowledge concerning mechanisms of action for an organism is
limited. Complete certainty can be thought of
as a special kind of uncertainty.
Knowledge Level. A computer system performing a specific task has various functional
levels. The lowest level is the device level, or
bit level. The highest level in a traditional
computer program is the program, or symbolic
level, and is understandable by most people
with programming experience. Symbols are
representations of real world objects. A KBS
introduces a new computational level above
the symbol level: the knowledge level. A
knowledge representation scheme can be considered to be a reduction of knowledge from
the knowledge level to the lower symbol level
Journal of Dairy Science Vol. 77, No. 12, 1994
(41). Most knowledge can easily be represented
in symbols; i.e., it can easily be transferred
from one person to another. Knowledge that
can easily be transferred into symbols is defined as symbolic knowledge. Subsymbolic
knowledge, however, is knowledge that cannot
be transferred easily in symbols; i.e., subsymbolic knowledge is difficult to explain to
others. The knowledge involved in pattern
recognition is regarded as subsymbolic knowledge. For example, in the analysis of milking
curves, the dairy f m advisor can recognize
problems in milk production immediately,
even when the data in the curve are not complete, explaining why a certain production
curve indicates a milk production problem
takes much more time. Knowledge involved in
curve interpretation can therefore be considered as subsymbolic knowledge.
Methods for Knowledge Representation
Various knowledge representation methods
have been developed for use in decision support systems. Some methods included and
adopted for public domain, shareware, or commercial software packages can be used to
facilitate development of decision support systems. Other methods are currently under development in artificial intelligence research
laboratories, and those methods are often
suited only for one subject matter domain. The
knowledge representation methods that have
product development software and that appear
to be promising for use with dairy science
applications are described in the following sections. These selected methods and their various
strengths and weaknesses for declarative and
procedural knowledge characteristics are summarized in Figure 1.
Production Rules. The best known and most
applied knowledge representation scheme is
the use of production systems. The procedural
knowledge in a production system is
represented by a set of rules by which the
domain procedural knowledge is incorporated:
a database of domain declarative knowledge
and an inference mechanism for applying the
rules to the database. The rules (if . . . then
rules) represent condition and action pairs. The
antecedent (if) of a rule is a condition for the
rule to be applicable, and the consequent (then)
of a rule is the action that results when the rule
3707
OUR INDUSTRY TODAY
Type of Knowledge and Knowledge Characteristics
Knowledge
Representation
Methodology
Declarative Knowledge
Complete
-
Operative Knowledge
~
h c o m p i e t e Certain
Uncertain
_
_
_
_
Generic * * ‘ - Heuristic
Certain ”*Uncertain Symbolic*Bubsymbolic
Production
Rules
Fuzzy
Logic
Conditional
Causal
Model
X
x
X
Bayesian
Belief
Network
Neural
Network
x
Simulation
x
x
knowledge characteristic on righl
x
= Methodology better for knowledge characteristic left
= Inappropriate UY orthe methodology
x
X
= Only appropriate for knowkdge characteristic on left
= Only appropriate for knowledge characteristic on right
= Equally appropriate for bMh knowledge characteristics
I
Figure 1. Summarization of strengths and weaknesses of various knowledge representation schemes.
is applied (35). Forward and backward reasoning are the most common inference mechanisms. Production rules are very efficient in
representing heuristic knowledge, but complete
declarative knowledge is needed to solve problems.
Several development tools, called shells,
based on production systems are commercially
available (18, 49). Because a variety of shells
are available, Meyer (39) described minimum
standards for the user interface for production
system development tools. Production systems
have been used in dairy decision support systems to analyze yearly economic performance
(6,53), evaluation of reproductive performance
(17, 34, 37), milk production performance (21,
67), and comparisons between desired
(planned) results with actual results (3, 66).
F u u y Logic. Using fuzzy set theory, variables can be associated with a membership
function that can take values between 0 and 1
to describe the meaning of the variable. The
basic features of fuzzy set theory can be defined as follows. If S is a set, and s is a
member of that set, a fuzzy subset 0 is then
defined as a membership function mF(s) that
describes the degree to which s belongs to F
(68). Using predefined membership functions,
it can be stated how true the statement “fever”
is when a temperature of 39.1’C is observed.
When the fuzzy set theory is applied in a
mathematical or computer system, it can be
referred to as fuzzy logic. Fuzzy logic is often
used with production rules to combine measures of possibility (32). These features make
fuzzy logic useful in situations in which
specification of the defining characteristics of
important but not directly observable features
is difficult.
A number of decision tool development
systems that utilize fuzzy logic are commercially available. Although fuzzy logic is
mainly applied in controller tasks, it is being
applied more frequently in decision support
systems (32). In Japan, fuzzy set theory is
applied to d a q farm economics (40).
Bayesian Belief Network. The theory of
Bayesian belief networks (BBN),also known
as causal probabilistic networks, is based on
Bayesian conditionalization. A BBN is,
qualitatively, a graph on which the nodes represent domain objects and the links between
Journal of Dairy Science Vol. 77, No. 12, 1994
3708
HOGEVEEN ET AL.
1,
B
A
C
B
I
Figure 2. Basic features of a Bayesian belief network.
Figure 3. Basic features of a conditional causal model.
nodes represent relations between these objects
(Figure 2). The knowledge is stated in a causal
direction: for example, diseases cause symptoms. Each node in a BBN has a number of
states, describing the possible values of the
node. Quantitatively, the relationships expressed by the links are represented by conditional probabilities (4, 9). In a BBN, the conditional probabilities for each node on its parents
(in this case P(BIA)) are put .in a probability
table. When the state of node B has been
observed, the conditional probability P(AIB)
can be calculated using the probability table.
A BBN can be created with the decision
support development tool, HUGIN (4) (Hugin
Expert A / S , Aalborg, Denmark). A BBN,
which is very useful in modeling uncertainty,
can also reason with incomplete knowledge. In
dairy science, BBN have been used to examine
the cow’s reproduction (48), to diagnose mastitis caused by environmental factors (l), and to
determine the blood group of Danish Jersey
cattle (47).
Conditional Causal Model. A conditional
causal model (CCM) consists of a set of nodes
that describe a domain. The nodes are connected by a set of unidirectional links,
representing a causal dependency of a node on
another node. The magnitude of the dependency may be influenced by one or more conditions (24, 54). The relationships in a CCM
may be qualitative and quantitative. The basic
elements of CCM are graphically represented
in Figure 3. Node B is causally dependent on
node A, a relationship represented by the
unidirectional arrow. Node C is a condition,
represented by a circle on an arrow. Forward
reasoning (simulation) and backward reasoning
(diagnosis) are possible with a CCM. A CCM
is very flexible and allows the generic
representation of complex procedural knowledge. Data or observations used for input need
to be complete, and a CCM allows no reasoning with uncertainty.
A CCM can be developed using the tool
CAMEL (causal modeling environment and
laboratory; Laboratory for Artificial Intelligence, Erasmus University, Rotterdam, The
Netherlands) (61). Using the graphical interface, models in a domain can be made. The
relationships between the nodes can then be
quantified with underlying functions, written
in the artificial intelligence programming language Common LISP (58).
Hogeveen et al. (25) employed CCM for the
diagnosis of herd mastitis problems. Schakenraad et al. (51, 52) used CCM to support
decisions regarding feed and grassland utilization on dairy farms.
Neural Network. A neural network, a model
consisting of layers of highly interconnected
processing units, can be trained to perform
classification tasks. Patterns of input and output are first presented to the model for training. The subsymbolic knowledge of a trained
model is implicitly stored in the weights of the
connections or arrows pointing toward and
away from the internal representation units
(Figure 4) (50). Various methods exist to train a
neural network; the most frequently used is
back-propagation with the generalized delta
rule (50). With back-propagation, the user defines the number of hidden layers and nodes in
each layer. Then, the model generates a first
output, based on random weights of the connections. This output is compared with the
desired output, and the difference between
model prediction and desired output is calcu-
Journal of Dairy Science Vol. 77, No. 12, 1994
OUR INDUSTRY TODAY
I l l
4
I
Internal
Representation
Units
3709
64). Simulation models require complete and
certain declarative knowledge as input. The
procedural knowledge can utilize measures of
certainty, in which case, the simulation model
is stochastic (13). Integration of heuristic
knowledge into simulation models can enhance
the performance of those models (3, 27, 45).
However, simulation models often provide a
solution that is precise but may not be overly
robust. A model-based reasoning system can
therefore be a good alternative (46).
DOMAIN SPECIFICATION IN DAIRY
FARM MANAGEMENT
Domain Classification In Mastltis
On a dairy farm, mastitis problems can
occur for individual cows and for herds. For
individual cows, a mastitis problem is a cow
with clinical or subclinical mastitis. A decision
must then be made about treatment of the cow.
Input Patterns
When the incidence of mastitis on a farm is
high, a mastitis problem for the herd exists.
Figure 4. A multilayer (inpuL output, and one hidden For mastitis for herds, a diagnosis must be
layer) neural network. Based on data from Rumelhart et al.
made and the possible causes of the mastitis
(50).
problem determined.
In general, decision making about solutions
to problems involves three stages: 1) problem
lated. The total squared sum of the calculated detection, 2) problem diagnosis, and 3) decidifferences is then returned into the model, and sion generation. When these stages are applied
the weights of the connections are changed to to a system for automated detection and diagminimize the error. This procedure is repeated nosis of mastitis, the resulting system contains
many times for all combinations of input and six subsystems (Figure 5).
output. The ultimate goal for the model is to
find a single set of weights that satisfies all the
pairs of input and output presented to it, which
is generalized to classify new data correctly.
Neural networks are good at pattern recognition; they require no assumptions on data
frequency or distribution, and, after training,
neural networks can perform classification
tasks with missing data (43). Thus, trained
neural networks can function with incomplete
declarative knowledge.
Simulation Models. Simulation models
make use of arithmetic instead of artificial
intelligence or reasoning, and they are often
not considered to be a knowledge representation method. However, simulation models represent knowledge in some ways (65), and, furFigure 5. Diagram of the modules in a decision support
thermore, they are widely applied in decision system for mastitis management. BBN = Bayesian belief
support systems for dairy management (12, 29, network; CCM = conditional causal model.
Journal of Dairy Science Vol. 77, No. 12, 1994
3710
HOGEVEEN ET AL.
Mastitis Detection. Mastitis detection on a
cow level is normally performed by the milker
during the udder preparation before milking.
Abnormal milk is a key indicator of clinical
mastitis. Diagnosis of subclinical mastitis is
more difficult. Even the definition of subclinical mastitis is not clear. An association exists
between electrical conductivity of milk and
mastitis (42). Electrical conductivity can be
measured in the milking cluster and is therefore suitable for use in on-line detection of
mastitis (36). Because much of the information
in electrical conductivity data is organized in
the data pattern, the knowledge used to detect
mastitis from electrical conductivity patterns is
highly subsymbolic. Therefore, a neural network is a good knowledge representation
methodology to use in the detection of mastitis
at the cow level (43). Initial results of using a
neural network for mastitis detection have
been described and seem promising (44).
Pathogen Diagnosis. To treat mastitis
properly, diagnosis of the pathogen causing the
mastitis is very helpful. Based on clinical examination of the cow, cow history, and herd
history, the veterinarian makes a likely diagnosis. The diagnostic reasoning process involves
uncertain knowledge (S), and the diagnosis is
stated in terms of a likelihood. Furthermore,
the declarative knowledge is mostly incomplete. These characteristics suggest that BBN
is an appropriate approach for knowledge
representation.
Therapy Selection. Once a veterinarian has
made a likely diagnosis of pathogen, a decision
is made concerning the proper treatment of
that cow. Treatment could be antibiotic therapy, but it might also be culling the cow. The
reasoning process for therapy advice is heuristic. For a certain pathogen, an antibiotic, alone
or in combination with another antibiotic, will
be advised. Advice on antibiotic treatment can
be extended with other treatment, such as early
drying off. Therefore, the most efficient
method by which to model the knowledge
involved in the therapy selection is production
rules. The decision about whether to cull a cow
when mastitis is diagnosed depends on the
future profitability of the specific cow compared with that of a replacement cow. Stochastic modeling can be used to perform the calculations of future profitability (26, 64).
Problem Detection. A mastitis problem for
the herd is normally detected by the farmer or,
Journal of Dairy Science Vol. 77, No. 12, 1994
when the farm is in a veterinary herd health
program, by the veterinarian. When a good
historical database concerning dairy herd
health on a farm is available, automated problem diagnosis can be performed (15). Although
exact guidelines for mastitis detection may be
difficult to provide, the mastitis incidence rate
or a change in incidence rate can indicate a
mastitis problem for the herd. When subclinical mastitis is taken into account, SCC can be
used to detect a mastitis problem for the herd
(16). Electrical conductivity might then be used
as a tool to estimate subclinical mastitis and to
screen for overall udder health (42). After the
herd history data have been preprocessed,
heuristic knowledge is used to determine
whether a farm can be considered to have a
mastitis problem. Therefore, production rules
can be used to model the knowledge involved
in the detection of mastitis problems for herds.
Causal Diagnosis. To diagnose the causes
of a herd mastitis problem, a combination of
specialized knowledge from various domains
is necessary. The causes of a mastitis problem
can, for instance, include a malfunctioning
milking machine, improper milking techniques, suboptimal housing, or a deficiency in
udder defense. Much of the knowledge involved in the herd level mastitis diagnosis can
be described as textbook or generic knowledge. The interrelationship between various
causes in the causal diagnosis is very complex.
Also, for the herd level diagnosis, uncertainty
does not play an important role. An appropriate knowledge representation method would
therefore be a CCM.
Advice. When causes for a herd mastitis
problem are found, advice must be generated
to eliminate the causes. Knowledge involved in
the advice is relatively straightforward text
knowledge on options to eliminate a cause.
When a CCM is used, information boxes with
specific advice to eliminate a cause can be
included at appropriate places in the system.
Domain Classification for Time of AI
Inefficient reproduction causes significant
losses in profitability of dairy herds (7). Inaccurate or inefficient detection of estrus is
thought to be the leading cause of these losses.
Detection of estrus is required to identify optimal time for AI. A number of estrus detection
371 1
OUR INDUSTRY TODAY
aids have been proposed, and their use has
been reviewed (20, 33). but none has been both
accurate in identification of estrus and efficient
in identifying all estrus periods. Use of an
expert system to identify estrus has been proposed (57), but no details on the operation of
that system were provided. A hybrid decision
support system that utilizes the knowledge
described could be constructed to identify optimal time for AI. Such a KBS must be able to
reason using knowledge from different
sources. The system consists of eight components (Figure 6), seven components concerned
with the interpretation of data from the various
data sources for estrus detection and an integrating module to interpret the knowledge
provided by the other modules in the system.
The characteristics of the components involved
in a KBS for optimalization of the AI time are
described herein.
Dates of Previous Estrus. The dates of
previous estrus are useful in predicting the date
of the next estrus or in c o n f i n g behavior
indicating estrus. Because many estrus periods
are unobserved, this kind of knowledge is
likely to be incomplete. Not all cows are accurately identified as being in estrus, so some
uncertainty is associated with this knowledge.
Fuzzy set theory is the knowledge representation method that matches these characteristics
of declarative knowledge.
Dates of Previous AI. The dates of previous
AI are related to dates of previous estrus.
However, because cows are not always inseminated, some dates of previous estrus have no
AI, and thus are a separate source of knowledge with different characteristics. The AI
dates are utilized for other purposes, such as
prediction of parturition dates and billing for
AI services. Consequently, dates of previous
AI are likely to be complete and certain because of their importance. Various methods of
knowledge representation could be used with
complete and certain knowledge, but fuzzy set
theory would provide some flexibility in
representing the normal variation in estrus cycle length that would be used to interpret
current knowledge using previous AI dates.
Dates of Previous
Dates of Previous
Veterinary
Observation
Estrus
[Fuzzy Set Theory]
AI
[Fuzzy Set Theory]
Palpation Results
[BBN]
of cows
Traditional Aids
Milk Progesterone
For Estrus
Detection
[Production Rules]
Concentration
[BBNI
Electronic Data
- Milk Production
- Feed Consumed
- Pedometer
Integrating Module
To Identify Time For AI
[Fuzzy Set Theory]
Figure 6. A theoretical diagram of a hybrid decision suppoa system to identify optimal time for AI. BBN = Bayesian
belief network.
Journal of Dairy Science Vol. 77, No. 12, 1994
37 12
HOGEVEEN ET AL.
Results of Veterinary Palpation. Results of
rectal examination of the cow reproductive
tract by a veterinarian are often used by the
dairy farmer to detect pregnancy, to predict
estrus, or to identify bovine ovarian dysfunction. The veterinarian usually makes infrequent
or irregular examinations of a cow; therefore,
those data tend to be incomplete. Research (38)
has shown that predictions of estrus or bovine
ovarian dysfunction are often inaccurate at
predicting estrus or bovine ovarian dysfunction. Although the accuracy of the results may
be uncertain, the veterinary assessments are
typically certain. The veterinarian's observations are the declarative part of the knowledge
and can be considered to be certain, but the
procedural part of the knowledge contains uncertainty. Thus, a knowledge representation
method such as a BBN would be very appropriate for most types of results of rectal
palpation.
Observations of Cows. Observation of cow
behavior by the dairy fanner is the primary
method for identification of estrus and determination of AI time, but the declarative knowledge is incomplete because cows are not observed 24 h/d. As with results from rectal
palpation, not all cows identified in estrus are
truly in estrus, and the procedural knowledge
would thus be uncertain. However, the producer is frequently certain of the observations.
A BBN would be appropriate for observations
of cows that contain incomplete and relatively
certain declarative knowledge.
Traditional Devices for Estrus Detection.
Various commercial devices for estrus detection are available, and they are used to varying
degrees on dairy farms (33). Typically, these
devices show many false positives, reflecting a
high degree of uncertainty in the procedural
component of the knowledge concerning
devices for detection of estrus. For herds in
which these devices are used, the declarative
component of the knowledge is frequently
complete (the device is used on all cows) and
certain (the device gives yes or no as an answer). Production rules could then be used as
knowledge representation method with these
devices for detection of estrus.
Progesterone Concentrations in Milk.
Progesterone concentrations in milk can be
used to determine whether the ovary has an
active corpus luteum and that knowledge can
Journal of Dairy Science Vol. 77, No. 12, 1994
be used to infer other knowledge (38). For
instance, high progesterone concentrations in
milk would be associated with days of very
low cow fertility, and AI would not be indicated. Data on progesterone in milk would be
relatively incomplete because milk samples are
infrequently analyzed for progesterone, but the
knowledge from the levels would be highly
certain, suggesting that a BBN would a proper
knowledge representation method.
Electronic Data. An increasing number of
electronically collected knowledge sources are
becoming available on dairy farms. Electronic
pedometers have been utilized to record the
activity of dairy cows; activity increases when
the cow is in estrus (33). Electronic mount
detectors have also been reported recently (14).
Milk production and feed consumed sometimes decreased during estrus (2). Typically,
the data associated with these electronic
sources are complete unless a sensor fails. The
declarative component of the knowledge is
certain. The certainty of the procedural component for the knowledge varies widely according to the knowledge source. A BBN is a
useful knowledge representation method for
this kind of electronically collected data.
Integration Module. The knowledge
provided by the components in the hybrid KBS
described must then provide knowledge to an
integration module (bottom of Figure 6). This
integration module would use fuzzy set theory
to combine information from the other modules and to predict a time for AI.
DISCUSSION
Knowledge representation is of great interest in research of artificial intelligence. A
complete review of knowledge representation
schemes is therefore beyond the scope of this
paper. The knowledge representation schemes
described herein include the various knowledge characteristics that tend to be public domain or available in commercial software
packages. Therefore, these methods can easily
be applied in development of decision support
systems for dairy farms.
When KBS are applied in such systems, use
of the proper knowledge representation scheme
is important to profit from the strengths while
avoiding the weaknesses of a single method.
Use of the proper representation scheme en-
OUR INDUSTRY TODAY
hances the efficiency of the system. In a larger
decision support system, a series of systems
can be applied to solve a problem efficiently.
The examples of mastitis and reproduction
management illustrate this concept. A similar
approach was also used by Serodes and
Rodriguez (56) in management of drinking water quality. Also, support systems for operational decision making can be a part of a
management support system for tactical decision makmg.
The knowledge-based methods described
herein are exciting tools in the development of
decision support systems and need to be added
to existing techniques to apply information
processing to obtain functional decision support systems that are accepted by the farming
community.
ACKNOWLEDGMENTS
This research has been made possible by the
SKBS (Foundation for Knowledge Systems).
The authors gratefully thank S. K. Andersen of
the University of Aalborg, Denmark, for his
advice in Bayesian Belief Networks.
REFERENCES
1 Agger, J. F., K. G . Olesen, F. V. Jensen. and S. K.
Andersen. 1990. Computer-aided decision making in
mastitis control: development of a causal probabilistic
network (CPN). Addendum in Proc. 2nd GIL-Symp.,
Bonn, Germany.
2 Allrich, R. D. 1993. Estrous behavior and detection in
cattle. Vet. Clin. North Am. Food Anim. Pract. 9:249.
3 Amir, I., J. Puech, and J. Granier. 1991. ISFARM an integrated system for farm management 1. Methodology. Agric. Syst. 35:455.
4Andersen. S. K., K. G. Olesen, F. V. Jensen. and F.
Jensen. 1989. HUGIN-a shell for building belief
universes for expert systems. Page 180 in Proc. loth
Int. Joint Conf. Artificial Intelligence. Morgan Kaufmann Publ.. San Mateo. CA.
5 Brand, A., H. Folkers, D. W. de Hoop, W.J.A.
Hanekamp. J. H.Portiek, G. J. Rooker. J. W. Seinhorst, L.H.P. Janssen, G.M.A. Verheijen, L.H.M.
Mathijssen, and F. Balfoort. 1986. Information model
on dairy farming. Res. Stn. Cattle, Sheep and Horse
Husbandry, Lelystad, The Netherlands.
6 B d e . D. S., and W.H.G.J. Hennen. 1989. The use of
expert systems for the interpretation of technical and
economic performance of dairy farms. Page 191 in
Expertensystemen in der Agranvirtschaft Entwicklung, Erfahrung. Perspektiven. G. Schiefer, ed. Proc.
ITMA-Fachtagung 'Expertensysteme'. Kiel, Germany.
7 Britt, J. H. 1985. Enhanced reproduction and its economic implications. J. Dairy S c i . 68:1585.
8 Chamberlain, A. T. 1992. Predictive models: application of new technologies to clinical decision making.
3713
Page 50 in Proc. loth SOC.Vet. Epidemiol. Prev. Med.
Mtg. Edinburgh, Scotland.
9 Charmak, E. 1991. Bayesian networks without tears.
AI Mag. 12(4):50.
10 Davis, R. 1987. Knowledge-based systems: the view
in 1986. Page 13 in AI in the 1980's and Beyond.
W.E.L. Grimson and R. S. Patil. ed. MIT Press,
Cambridge, MA.
11 Davis, R., H.Shrobe. and P. Szolovits. 1993. What is
a knowledge representation? AI Mag. 14(1):17.
12DeLorenz0, M. A., T. H.Spreen, G. R. Bryan, D. K.
Beede, and J.A.M. van Arendonk. 1992. Optimizing
model: insemination, replacement, seasonal production, and cash flow. J. Dairy Sci. 75:885.
13 Dijkhuizen. A. A. 1988. Modelling to support health
programs in modem livestock farming. Neth. J. Agric.
Sci. 36:35.
14 Doh, H.. A. Yamada, S. Tsuda, T. Sumikawa, and S .
Entsu. 1993. Technical note: a pressure-sensitive sensor for measuring 'the characteristics of standing
mounts of cattle. J. Anim. Sci. 71:369.
15Dohoo. I. R. 1992. Dairy APHIN-an
information
service for the dairy industry in Prince Edward Island,
Canada. Prev. Vet. Med. 12:259.
16Dohoo. 1. R., and K. Leslie. 1991. Evaluation of
changes in somatic cell counts as indicators of new
intramammary infections. Prev. Vet. Med. 10:225.
17 Domecq, I. J., R. L. Nebel, M. L. McGilliard, and A.
T. Pasquino. 1991. Expert system for evaluation of
reproductive performance and management. J . Dairy
Sci. 74:3446.
18Engel. B. A,, D. D. Jones, J. R. Wright, and S .
Benabdallah. 1991. Selection of an expert system
development shell. AI Appl. Nat. Resource Manag.
5( 1):15.
19Evans. M., R. Mondor, and D. Flaten. 1989. Expert
systems and farm management. Can. J. Agric. Econ.
37:639.
20 Foote, R. H. 1975. Estrus detection and estrus detection aids. J. Dairy Sci. 58:248.
21 Fourdraine, R. H., M. A. Tomaszewski. and T. J.
Cannon. 1992. Dairy herd lactation expert system, a
program to analyze and evaluate lactation curves.
Page 331 in Proc. Int. Symp. Prospects Automatic
Milking. A. H. Ipema, A. C. Lippus, J.H.M. Metz, and
W. Rossing, ed. Eur. Assoc. Anim. Prod. Publ. No.
65. Pudoc Sci. Publ., Wageningen, The Netherlands.
22 Harsh, S. B. 1988. Artificial intelligence-methods,
tools and importance of knowledge acquisition. Page
176 in Proc. Deutsche Landwirtschafi Gesellschaft
Symp.. Frankfurt, Germany.
23 Hogeveen, H., E. N. Noordhuizen-Stassen, J. F.
Schreinemakers, and A. Brand. 1991. Development of
an integrated knowledge-based system for management support on dairy farms. J. Dairy Sci. 74:4377.
24Hogeveen. H., J. F. Schreinemakers, and E. N.
Noordhuizen-Stassen. 1992. Applications of conditional causality in an integrated knowledge-based system for dairy farms. AI Appl. Nat. Resources Agric.
Environ. Sci. 6(3):5.
25Hogeveen. H., J. van Vliet. E. N. NoordhuizenStassen, and A. Brand. 1992. Development of a
knowledge-based system describing the relations between mastitis and milking machines. Page 385 in
Roc. Int. Symp. Prospects Automatic Milking. A. H.
Ipema, A. C. Lippus, J.H.M. Metz. and W. Rossing,
ed. Eur. Assoc. Anim. Prod. Publ. No. 65. Pudoc Sci.
Publ., Wageningen, The Netherlands.
Journal of Daily Science Vol. 77, No. 12, 1994
37 14
HOGEVEEN ET AL.
26 Houben, E.H.P., R.B.M. Huime, and A. A. Dijkhuizen. 1991. An integrated dynamic programming
and expert system approach in livestock management.
Page V1.2.1 in Proc. 26th Sem. E o n . Artificial Intelligence Agric., Grignon, France.
27 Huirne, R.B.M.. A. A. Dijkhuizen, A. Pijpers, J.H.M.
Verheijden, and P. van Gulick. 1991. An economic
expert system on the personal computer to support
sow replacement decisions. Prev. Vet. Med 11:79.
28 Jackson, P. 1990. Introduction to Expert Systems. 2nd
ed. Addison-Wesley Publ. Co., Reading, MA.
29 Jalvingh, A. W. 1992. The possible role of existing
models in on-farm decision support in dairy cattle and
swine production. Livest. Prod. Sci. 31:351.
30 Jones, L. R., and D. E. Dill. 1992. Introduction to
advanced computer technologies and terminologies.
Page 1 in Proc. Conf. Adv. Comp. Appl. Anim.
Agric., Dallas, TX.
31 Jones, P. 1989. Agricultural applications of expert
systems concepts. Agric. Syst. 31:3.
32 Kosko. B., and S. Isaka. 1993. Fuzzy logic. Sci. Am.
269( 1):62.
33Lehrer. A. R., G. S. Lewis, and E. Aizinbud. 1992.
Oestrus detection in cattle: recent developments.
Anim. Reprod. Sci. 28:355.
34Levins. R. A., and M. A. Varner. 1987. An expert
diagnostic aid for reproductive problems in dairy cattle. Comp. Electron. Agric. 2:47.
35 Luger, G. F.. and W. A. Stubblefield. 1989. Artificial
Intelligence and the Design of Expert Systems.
Benjamin/Cummings Publ. Co., Redwood City, CA.
36 Maatje, K., P.J.M. Huijsmans, W. Rossing, and P. H.
Hogewerf. 1992. The efficacy of in-line measurement
of quarter milk electrical conductivity, milk yield and
milk temperature for the detection of clinical and
subclinical mastitis. Livest. Prod. Sci. 30:239.
McKay, B., S. McCallum, and R. S. Moms. 1988. An
expert system program for diagnosing reproductive
problems in seasonal dairy herds. Page 480 in Proc.
5th Int. Symp. Vet. Epidemiol. Econ., Copenhagen,
Denmark.
McLeod, B. J., and M. E. Williams. 1991. Incidence
of ovarian dysfunction in post partum dairy cows and
the effectiveness of its clinical diagnosis and treatment. Vet. Rec. 128:121.
39 Meyer, C. R. 1990. Minimum user-interface standards
and software for agricultural expert systems. Agron. J.
82647.
40 Nagaki, M. 1992. Computer-aided dairy farm management decision-making in Japan-experiences and ongoing efforts in software development. Page 111 in
Proc. 4th Int. Congr. Comp Technol. Agric. Paris Versailles, France.
41 Newell, A. 1981. The knowledge level. AI Mag. 22):
1.
42 Nielen, M., H. A. DeLuyker, Y. H Schukken, and A.
Brand. 1992. Electrical conductivity of milkmeasurement, modifiers, and meta analysis of mastitis
detection performance. J. Dairy Sci. 75:606.
43 Nielen, M., H. Hogeveen, Y. H. Schukken, H. A.
DeLuyker, and J. F. Schreinemakers. 1991. Using a
connectionist model (neural network) to analyse online milking parlour data to detect mastitis. Page 258
in Proc. 6 Int. Symp. Vet. Epidemiol. Econ.. Ottawa,
ON. Canada.
Journal of Dairy Science Vol. 77. No. 12, 1994
44 Nielen, M.. M. H. Spigt. and K. Maatje. 1992. Detecting mastitis with a neural network using electrical
conductivity data. Page 370 in Proc. Int. Symp.
Prospects Automatic Milking. A. H. Ipema, A. C.
Lippus, J.H.M. Metz, and W. Rossing, ed. Eur. Assoc.
Anim. Prod. Publ. No. 65, Pudoc Sci. Publ., Wageningen, The Netherlands.
45 Plant, R. E., R. D. Horrocks, D. W. Grimes, and L. J.
Zelinski. 1992. CALEX/Cotton--an integrated expert
system application for irrigation scheduling. Trans.
Am. SOC.Agric. Eng. 35:1833.
46Plant. R. E., and R. S. Loomis. 1991. Model-based
reasoning for agricultural expert systems. AI Appl.
Nat. Resource Manag. 5(4):17.
47Rasmussen. L. K. 1992. A causal probabilistic network for blood group determination of Danish Jersey
cattle. Dina Res. Rep. No. 3, Danish Inst. Anim. Sci.,
Tjele, Denmark.
48 Rasmussen, L. K., I. Thysen, and K. M. Pedersen.
1990. An application of causal probabilistic networks
to examine reproduction of dairy cows. Page 59 in
Proc. Workshop Expert Syst. Agric. Res., Ebeltoft,
Danish Res. Sew. Plant Soil Sci., Natl. Inst. Anim.
Sci., Res. Ctr., Foulum Denmark.
49 Richer, M. H. 1986. An evaluation of expert system
development tools. Expert Syst. 3:166.
50 Rumelhart, D. E.. G. E. Hinton. and R. J. Williams.
1986. Learning internal representations by error
propagation. Page 318 in Parallel Distributed Processing: Explorations in the Micro-Structure of Cognition.
Vol. 1. D. E. Rumelhart and J. L. McClelland, ed.
MIT Press, Cambridge, MA.
51 Schakenraad, M.H.W. 1993. A multi level conditional
causal model for planning and diagnosis of farm
performance. Workshop Notes Workshop AI Agric.
Nat. Resources Environ. Sci., Chambery, France.
52 Schakenraad, M.H.W., W.H.G.J. Hennen, and D. W.
de Hoop. 1991. Development of a knowledge based
system for the evaluation of the feed and grassland
management on dairy farms. Page VI.3.1. in Proc.
26th Sem. Econ. Artificial Intelligence Agric., Grignon, France.
53 Schmisseur, E., and M. J. Gamroth. 1993. DXMAS:
an expert system program providing management advice to dairy operators. J. Dairy Sci. 76:2039.
54Schreinemakers, J. F. 1991. Pattern recognition and
symbolic approaches to diagnosis. Ph.D. Diss., Erasmus Univ., Rotterdam, The Netherlands.
55Schreinemakers, J. F., M. L. Vos, A. Brand, D. S.
Bke, and J.H.M. Verheijden. 1988. The introduction
of expert systems in animal husbandry. Vet. Q. 10:
281.
56Serodes, J. B., and M. J. Rodriguez. 1993. Using
expert systems and neural networks to manage drinking water quality in distribution systems. AI Appl.
Nat. Resources, Agric. Environ. Sci. 7(1):44.
57 Spahr, S. L., L. R. Jones, and D. E. Dill. 1988. Expert
systems-their use in dairy herd management. J.
Dairy Sci. 71:879.
58 Steele, G. L., Jr. 1990. Common LISP: the language.
2nd ed. Digital Press, Bedford, MA.
59 Stock. M. 1992. Knowledge engineering and knowledge acquisition. Page 85 in Proc.Conf. Adv. Comp.
Appl. Anim. Agric., Dallas, TX.
OUR INDUSTRY TODAY
60Szolovits. P. 1987. Expert systems tools and techniques: past, present and future. Page 43 in AI in the
1980s and Beyond. W.E.L. Grimson and R. S. Patil,
ed. MIT Press, Cambridge, MA.
61 Tepp, D. M., and J. F. Schreinemakers. 1991.
CAMEL: causal model environment and laboratory.
Manag. Rep. Ser. 97. Erasmus Univ., Rotterdam, The
Netherlands.
62 Tomaszewski, M. A. 1992. Using advanced computer
technologies to increase extension effectiveness. J.
Dairy Sci. 753242.
63 Tucker, T., R. Shannon, and R. Lovellette. 1990. A
comparison of the development methodologies between conventional systems and expert systems. AI
Appl. Nat. Resource Manag. 4(2):27.
64 Van Arendonk, J.A.M. 1984. Studies on the replacement policies in dairy cattle. I. Evaluation of tech-
3715
niques to determine the optimum time for replacement
and to rank cows on future profitability. Z. Tierz.
Zuechtungsbiol. 101:330.
65 Wagner, P. 1993. Techniques of representing knowledge in knowledge-based systems. Agric. Syst. 4153.
66 Wagner, P.. and F. Kuhlmann. 1991. Concept and
implementation of an integrated decision support system (IDS) for capital-intensive farming. Agric. &on.
5:287.
67 Whittaker, A. D., M. A. Tomaszewski, J. F. Taylor, R.
Fourdraine, C. J. van Overveld, and R. G. Schepers.
1989. Dairy herd nutritional analysis using knowledge
systems techniques. Agric. Syst. 31:83.
68Zadeh. L. A. 1983. Commonsense knowledge
representation based on fuzzy logic. Computer 16(10):
61.
Journal of Dairy Science Vol. 77, No. 12, 1994