Professional Documents
Culture Documents
Articles
The Process
Specification
Language (PSL)
Theory and Applications
Michael Grüninger and Christopher Menzel
■ The PROCESS SPECIFICATION LANGUAGE (PSL) has been creased and the information has become more
designed to facilitate correct and complete ex- complex, it has been more difficult for soft-
change of process information among manufac- ware developers to provide translators between
turing systems, such as scheduling, process model-
every pair of applications that must cooperate.
ing, process planning, production planning,
What is needed is some way of explicitly spec-
simulation, project management, work flow, and
business-process reengineering. We give an ifying the terminology of the applications in
overview of the theories within the PSL ontology, an unambiguous fashion.
discuss some of the design principles for the ontol- The PROCESS SPECIFICATION LANGUAGE (PSL)
ogy, and finish with examples of process specifica- (Menzel and Gruninger 2001; Schlenoff,
tions that are based on the ontology. Gruninger, and Ciocoiu 1999) has been de-
signed to facilitate correct and complete ex-
change of process information among manu-
A
s the use of information technology in facturing systems. Included in these
manufacturing operations has matured, applications are scheduling, process modeling,
the need to integrate software applica- process planning, production planning, simu-
tions has become increasingly important. lation, project management, work flow, and
However, interoperability among these manu- business-process reengineering. We give an
facturing applications is hindered because the overview of the theories within the PSL ontol-
applications use different terminology and rep- ogy, discuss some of the design principles for
resentations of the domain. These problems the ontology, and finish with examples of
arise most acutely for systems that must man- process specifications that are based on the on-
age the heterogeneity inherent in various do- tology.
mains and integrate models of different do-
mains into coherent frameworks (figure 1). For
example, such integration occurs in business-
Architecture of PSL
process reengineering, where enterprise mod- The PSL ontology is organized into PSL-CORE and
els integrate processes, organizations, goals, a partially ordered set of extensions. All axioms
and customers. Even when applications use the are first-order sentences and are written in the
same terminology, they often associate differ- KNOWLEDGE INTERCHANGE FORMAT (KIF).
ent semantics with the terms. This clash over There are two types of extensions within PSL:
the meaning of the terms prevents the seam- (1) core theories and (2) definitional exten-
less exchange of information among the appli- sions. Core theories introduce and axiomatize
cations. Typically, point-to-point translation new relations and functions that are primitive.
programs are written to enable communica- All terminology introduced in a definitional
tion from one specific application to another. extension has conservative definitions using
However, as the number of applications has in- the terminology of the core theories. Thus, def-
Copyright © 2003, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2003 / $2.00 FALL 2003 63
Articles
Scheduler
Throughput, Process Plans
initional extensions add no new expressive come necessary to describe all intuitions about
power to PSL-CORE. manufacturing processes. To supplement the
concepts of PSL-CORE, the ontology includes a
Core Theories set of extensions that introduce new terminol-
All core theories within the ontology are con- ogy. Any PSL extension provides the logical ex-
sistent extensions of PSL-CORE, although not all pressiveness to axiomatize intuitions involving
extensions need be mutually consistent. Also, concepts that are not explicitly specified in PSL-
the core theories need not be conservative ex- CORE. All extensions within PSL are consistent
tensions of other core theories. The relation- extensions of PSL-CORE and can be consistent
ships among the core theories in the PSL ontol- extensions of other PSL extensions. However,
ogy are depicted in figure 2. not all extensions within PSL need be mutually
PSL Core The purpose of PSL-CORE is to axiom- consistent. Also, the core theories need not be
atize a set of intuitive semantic primitives that conservative extensions of other core theories.
is adequate for describing the fundamental A particular set of theories is grouped togeth-
concepts of manufacturing processes. Conse- er to form the outer core; this distinction is
quently, this characterization of basic processes pragmatic because in practice, these theories
makes few assumptions about their nature be- have been necessary for axiomatizing all other
yond what is needed for describing these concepts in the PSL ontology (figure 2).
processes, and the core is, therefore, rather Occurrence Trees An occurrence tree is the set
weak in terms of logical expressiveness. Specif- of all discrete sequences of activity occur-
ically, the core ontology consists of four dis- rences. They are isomorphic to substructures of
joint classes: (1) activities, (2) activity occur- the situation tree from situation calculus (Mc-
rences, (3) time points, and (4) objects. Carthy and Hayes 1969; Pinto 1994; Reiter
Activities can have zero or more occurrences, 1991), the primary difference being that rather
activity occurrences begin and end at time than a unique initial situation, each occur-
points, and time points constitute a linearly or- rence tree has a unique initial activity occur-
dered set with end points at infinity. Objects are rence. As in the situation calculus, the poss re-
simply those elements that are not activities, lation is introduced to allow the statement of
occurrences, or time points. constraints on activity occurrences within the
PSL-CORE is not strong enough to provide de- occurrence tree. Because the occurrence trees
finitions of the many auxiliary notions that be- include sequences that modelers of a domain
64 AI MAGAZINE
Articles
Activity
Occurrences
Complex Activites
PSL-Core
will consider impossible, the poss relation activities have preconditions (fluents that must
“prunes” away branches from the occurrences hold before an occurrence) and effects (fluents
tree that correspond to such impossible activi- that always hold after an occurrence).
ty occurrences. Subactivities This core theory axiomatizes
It should be noted that the occurrence tree intuitions about subactivities. The only con-
is not the structure that represents the occur- straint imposed within this theory is that the
rences of subactivities of an activity. The oc- subactivity relation is isomorphic to a discrete
currence tree does not represent a particular partial ordering. Other core theories impose
occurrence of an activity but, rather, all possi- additional constraints.
ble occurrences of all activities in the domain. Atomic Activities The core theory of atomic
Discrete States The discrete-states core theo- activities axiomatizes intuitions about the con-
ry introduces the notion of state (fluents). Flu- current aggregation of primitive activities. This
ents are changed only by the occurrence of ac- concurrent aggregation is represented by the
tivities, and fluents do not change during the occurrence of concurrent activities rather than
occurrence of primitive activities. In addition, concurrent-activity occurrences.
FALL 2003 65
Articles
66 AI MAGAZINE
Articles
FALL 2003 67
Articles
The first aspect of this approach is to identify states of affairs, then we intuitively have an in-
the primary intuitions in some domain. With- correct set of structures.
in PSL, for example, we have intuitions about Once we specify the class of structures, we
concepts such as activity, activity occurrences, can formally evaluate an axiomatic theory
and time points. These intuitions also restrict with respect to this specification. In particular,
the scope of the axiomatic theories, and they we want to prove two fundamental properties:
serve as informal requirements that get formal- (1) satisfiability, where every structure in the
ly specified in the classes of structures and later class is a model of the axiomatic theory, and (2)
axiomatized in the theory itself. axiomatizability, where every model of the ax-
The objective of the second aspect of the iomatic theory is isomorphic to some structure
methodology is to identify each concept with in the class.
an element of some mathematical structure. In Strictly speaking, we only need to show that
particular, given a nonlogical lexicon in some a model exists to demonstrate that a theory is
language, the specified structures are isomor- satisfiable. However, in the axiomatization of
phic to the extensions of the relations, func- domain theories, we need a complete charac-
tions, and constants denoted by the predicate terization of the possible models. For example,
symbols, function symbols, and constant sym- because we are considering the domain of ac-
bols of the lexicon. The class of structures cor- tivities, occurrences, and time points to show
responding to the intuitions of the ontology that a theory is satisfiable, we need only specify
will be defined either by specifying some class an occurrence of an activity that, together with
of algebraic or combinatoric structures or ex- the axioms, is satisfied by some structure. The
tending classes of structures defined for other problem with this approach is that we run the
theories within the ontology. Examples of risk of having demonstrated satisfiability only
structures include graphs, linear orderings, par- for some restricted class of activities. For exam-
tial orderings, groups, fields, and vector spaces. ple, a theory of activities that supports sched-
This relationship between the intuitions and uling can be shown to be consistent by con-
the structures is, of course, informal, but we structing a satisfying interpretation, but the
can consider the domain intuitions as interpretation might require that resources
providing a physical interpretation of the cannot be shared by multiple activities, or it
structures. In this sense, we can adopt an ex- might require all activities to be deterministic.
perimental or empirical approach to the evalu- Although such a model might be adequate for
ation of the class of intended structures in such activities, it would in no way be general
which we attempt to falsify these structures. If enough for our purposes. We want to propose
we can find some objects or behavior within a comprehensive theory of activities, so we
the domain that do not correspond to an in- need to explicitly characterize the classes of ac-
tended structure, then we have provided a tivities, time points, objects, and other as-
counterexample to the class of structures. In re- sumptions that are guaranteed to be satisfied
sponse, we can either redefine the scope of the by the specified structures.
class of structures (that is, we do not include The purpose of the axiomatizability theorem
the behavior within the characterization of the is to demonstrate that any unintended models
structures), or we can modify the definition of of the theory do not work, that is, any models
the class of structures so that they capture the that are not specified in the class of structures.
new behavior. With the interoperability hypothesis, we do
For example, physicists use various classes of not need to restrict ourselves to elementary
differential equations to model different phe- classes of structures when we are axiomatizing
nomena. However, they do not use ordinary an ontology. Because the applications are
linear differential equations to model heat dif- equivalent to first-order inference engines,
fusion, and they do not use second-order par- they cannot distinguish between structures
tial differential equations to model the kine- that are elementarily equivalent. Thus, the un-
matics of springs. If we want to model some intended models are only those that are not el-
phenomena using a class of differential equa- ementarily equivalent to any model in the
tions, we can use the equations to predict be- class of structures.
havior of the physical system; if the predic-
tions are falsified by observations, then we The Role of Definitional Extensions
have an incorrect set of equations. Similarly, in The terminology within the definitional exten-
our case, we can use some class of structures to sions intuitively corresponds to classes of activ-
predict behavior or characterize states of af- ities and objects. Within the PSL ontology, the
fairs; if there is no physical scenario in the do- terminology arises from the classification of
main that corresponds to these behaviors or the models of the core theories with respect to
68 AI MAGAZINE
Articles
sets of invariants. Invariants are properties of ful. However, we find Petri nets undesirable as
models that are preserved by isomorphism. A a formal foundation for process modeling for
set of invariants is complete for a class of struc- at least two reasons: First, there is still no stan-
tures if and only if it can be used to classify the dard, widely agreed-on semantics for Petri nets,
structures to isomorphism. For example, a fi- and those semantic systems that exist are high-
nite abelian group can be classified to isomor- ly complex and require a sophisticated knowl-
phism by the subgroups whose orders are fac- edge of certain areas of mathematics. (The
tors of the group’s order. In general, it is not most common approach to providing a seman-
possible to formulate a complete set of invari- tics for Petri nets is to map the apparatus into
ants; for example, there is no known set of in- linear logic and then exploit one of several se-
variants that can be used to classify graphs to mantic theories for linear logic; see, for exam-
isomorphism. However, even without a com- ple, Marti-Oliet and Meseguer [1991]). Second,
plete set, invariants can still be used to provide Petri nets do not provide any sort of axiomatic
a classification of the models of a core theory theory of processes. It is therefore rather diffi-
in PSL, which leads to the classification criteri- cult to gain any insight, from Petri nets alone,
on and the definitional extension criterion: into the nature of the things that process mod-
Classification Criterion els are about and, hence, difficult to see how it
The set of models for the core theories of can serve as a basis for process-model sharing
PSL is partitioned into equivalence classes and integration.
defined with respect to the set of invari- By contrast, PSL scores well on both these
ants of the models. counts. The language of PSL has a rigorous se-
mantics that draws on familiar model-theoretic
Definitional Extension Criterion
and algebraic structures. This semantics, in
Each equivalence class in the classification
turn, is fully captured in a complete set of ax-
of PSL models can be axiomatized within a
ioms. We illustrate how it can be used as a
definitional extension of PSL.
foundation for process modeling with a simple
In particular, each definitional extension in example.
the PSL ontology is associated with a unique in- In general, business and engineering
variant; the different classes of activities or ob- processes are described at the type level: A
jects that are defined in an extension corre- process model characterizes a certain general, re-
spond to different properties of the invariant. peatable process structure. This structure, in
turn, might admit to many instances that, de-
PSL in Action: A Foundation for pending on how constrained the structure is,
might differ considerably from one another. A
Process Modeling robust foundation for process modeling, there-
In this section, we present simple examples to fore, should be able to characterize both the
illustrate one of the uses that PSL can be put to, general process structure described by a model
namely, as a foundation for the semantics and, as well as the class of possible instances of that
hence, integration, of business-process models. structure. Moreover, such a foundation must
Processes are patterns of activities. Process clearly be able to represent the constraints that
modeling is the linguistic, diagrammatic, or nu- a process model places on something’s count-
meric representation of such patterns. Process ing as an instance of the process, the con-
models, in these various forms, are ubiquitous straints, as we might say, on process realiza-
in industry: There is a plethora of business and tion.
engineering applications—work flow, schedul- A typical process is best thought of informal-
ing, discrete-event simulation, process plan- ly as a structured collection of activities that
ning, business-process modeling, and are related to one another in a manner that re-
others—that are designed explicitly for the flects the process flow and temporal relations
construction of process models of various sorts. that can appear in any given occurrence of the
The vision of enterprise integration, therefore, process. For example, consider the painting
will be realized only if it is possible to integrate process depicted in figure 4 (we use the graphic
business-process models. It is somewhat scan- notation of the IDEF3 process-description cap-
dalous that so little work on this rather practi- ture method to illustrate the intuitive process).
cal issue has been done. This diagram depicts a general process that
As is widely recognized, process-model inte- must begin with an occurrence of Paint Widget
gration will be possible only if there is a com- (represented by the Paint Widget box with no
mon semantics of process information to draw predecessor), followed by an occurrence of Test
on. Among potential semantic frameworks, the Coverage. At this point, depending on the out-
theory of Petri nets is perhaps the most power- come of the test, an occurrence of the process
FALL 2003 69
Articles
can either loop back to another occurrence of known as D’s preconditions and postconditions,
Paint Widget (wherein our current widget is re- respectively.
painted) or continue to have the widget dried. ARDs correspond roughly to IDEF3 boxes, as
Thus, there are, in principle, infinitely many seen in figures 4 and 5. Thus, in the context of
possible ways this single process can be instan- a process model, an ARD represents the struc-
tiated by a particular series of activity occur- tural role that the indicated activity plays in
rences, depending on how many times such a the process represented by the model. An ARD
series loops back to produce another occur- has both name (:name) and identifier (:id)
rence of the Paint Widget activity. fields because the same activity can play differ-
The PSL ontology axiomatizes the classes of ent roles in the same process. In such cases, we
activities and resources that are used when typically have two or more distinct declara-
defining a process. However, when using PSL, tions with the same activity name but with dis-
software applications are not exchanging defi- tinct activity identifiers because it is the iden-
nitions of classes of activities; rather, they are tifiers that indicate the distinct structural roles
exchanging sentences that are satisfied by ac- being played by the activity in the overarching
tivities that belong to these classes. Such sen- process. The successor field will contain the
tences are known as process specifications, and identifiers of other ARDs (or possibly the same
they include preconditions and effects, tempo- ARD) in the model and the precondition and
ral constraints on occurrences, and ordering postcondition PSL sentences that express condi-
constraints on subactivity occurrences. In the tions that must hold before and after an occur-
remainder of this article, we present several ex- rence of the given activity—in the indicated
amples of processes and a simplified syntax role—in a realization of the model.
that can be used for process specifications. In addition to activity constraints, one also
The first step is the notion of an activity role has to be able to express information about the
declaration (ARD), characterized generally as objects that participate in the activity occur-
follows: rences that jointly realize the model. Such in-
(define-activity-role formation is often relegated to text in a graphic
:id <number> model but is just as critical as the process-struc-
:name <string> ture information represented explicitly by the
:successors <number>* boxes and arrows of IDEF3. Hence, we introduce
:preconditions <PSL sentence>* a similar mechanism—object declarations—for
:postconditions <PSL sentence>*) introducing the participating objects into
process models:
The value of the :id field in an ARD D is
known as its role identifier, and the value of the (define-object
:name field of D is its activity name. The values :name <KIF constant>
of the :subactivity-successor field are known as :constraints <PSL sentence>*)
D’s successor identifiers, and the values of the The :name field of an object declaration, of
:preconditions and :postconditions fields are course, specifies the constant to be introduced,
70 AI MAGAZINE
Articles
and the :axioms field consists of PSL sentences cedure numerous times, as with the widget in
that characterize the indicated object. Identi- the previous process, but many objects of the
fiers are unnecessary because the same object same sort undergoing the same procedure, as
does not play different structural roles in a in a typical manufacturing process.
process the way that activities do. For example, the IDEF3 diagram in figure 5
Given this apparatus then, we can capture contains both sorts of looping. In this process,
both the structural information indicated by a widget is painted until coverage is adequate
the earlier IDEF3 diagram as well as implicit con- and then queued, at which point either anoth-
tent about participating objects. Note that a er widget is painted or, if the queue is full, all
general background ontology characterizing the queued widgets are dried en masse. Thus,
the relevant properties and relations in this the first loop, as in the first example, “carries”
model (Widget, In, Paint_Coverage) is being as- a single widget back to undergo an earlier ac-
sumed (figure 6). tivity, whereas the second indicates the begin-
Note that the preconditions and postcondi- ning of a new paint job with a new widget;
tions all contain a free activity occurrence vari- there is, so to speak, no “object flow” in the
able ?occ. Say that an activity occurrence e sat- second loop.
isfies an ARD if e is an occurrence of the activity The flexibility of variables in the PSL lan-
named in the ARD, and the preconditions and guage enables us to capture this semantics sim-
postconditions are true, relative to some vari- ply and easily. Their meanings, unlike ordinary
able assignment that assigns e as the value of names, can shift, and we can use this fact to al-
the occurrence variable (?occ is the only vari- low them to refer to different widgets in differ-
able in the previous ARDs). Because of the pres- ent events in a process realization. Like ordi-
ence of the occurrence variable, many different nary names, however, we can put constraints
occurrences of the activity can satisfy the same on the values of these variables. Following a
ARD, which is critical because looping can lead similar notion in situation theory (Barwise and
to a situation where the same ARD is satisfied Etchemendy 1987; Devlin 1991), we refer to
by many different occurrences—as happens in such constrained variables (and occasionally,
the case of our example if the same widget is re- ambiguously, their values) as parameters, and
painted because of inadequate coverage. we introduce a concomitant declaration tem-
A collection of activity occurrences can be plate:
said to realize a process model M if (1) the tem-
(define-parameter
poral-ordering overoccurrences in the collec-
tion can be mapped into the ordering deter- :variable <KIF variable>
mined by the successor fields of the ARDs in M :constraints <PSL sentence>*)
in a structure-preserving way (that is, homo- We now say that an occurrence e satisfies an
morphically) and (2) each activity occurrence in ARD D if it satisfies it in the previous sense; in
the collection satisfies the ARD to which it is addition, for any parameter V occurring in the
mapped. Given this definition, we see that any preconditions or postconditions of D, there is
series of activity occurrences in which a widget an object a participating in e such that the pa-
is painted and then, if necessary, repeatedly re- rameter’s constraints are true when V is as-
painted until its coverage is adequate and then signed a as its value.
dried will realize the process model. Armed with this construct, we can capture
This example, of course, is rather simplistic. the detailed semantics of the process indicat-
In particular, most iterative processes involve ed in figure 5 by adding the declarations in
not simply the same object undergoing a pro- figure 7.
FALL 2003 71
Articles
(define-object
:name widget
:constraints (Widget widget))
(define-object
:name painter
:constraints (Paint_Sprayer painter))
(define-object
:name oven
:constraints (Oven oven))
(define-activity-role
:id Act-1
:name Paint_Widget
:successors 2
:preconditions
(or (not (Painted widget (beginof ?occ)))
(not (Adequate (Paint_Coverage widget (beginof ?occ)))))
:postconditions
(Painted widget (endof ?occ)))
(define-activity-role
:id Act-2
:name Test_Coverage
:successors 1 3
:preconditions (Painted widget (beginof ?occ))
:postconditions (Adequate (Paint_Coverage widget) (endof ?occ)))
(define-activity-role
:id Act-3
:name Dry_Widget
:successors
:preconditions (Adequate (Paint_Coverage widget) (beginof ?occ))
:postconditions (Dry widget (endof ?occ))
Figure 6. Process Declarations for the Point, Test, and Dry Process.
72 AI MAGAZINE
Articles
(define-parameter
:variable ?w
:constraints (Widget ?w))
(define-activity-role
:id Act-1
:name Paint_Widget
:successors 2
:preconditions
(or (not (Painted ?w (beginof ?occ)))
(not (Adequate (Paint_Coverage ?w (beginof ?occ)))))
:postconditions
(Painted widget (endof ?occ)))
(define-activity-role
:id Act-2
:name Test_Coverage
:successors 1 3
:preconditions (Painted ?w (beginof ?occ))
:postconditions (Adequate (Paint_Coverage widget) (endof ?occ)))
(define-activity-role
:id Act-3
:name Queue_Widget
:successors 1 4
:preconditions
(Adequate (Paint_Coverage ?w (beginof ?occ)))
:postconditions
(Painted widget (endof ?occ)))
(define-activity-role
:id Act-4
:name Dry_Widget
:successors
:preconditions (Adequate (Paint_Coverage ?w) (beginof ?occ))
:postconditions (Dry ?w (endof ?occ))
Figure 7. Process Declarations for the Point, Test, Queue, and Dry Process.
Moreover, the modular organization of PSL en- Dennet, D. C. 1987. The Intentional Stance. Cam-
ables the flexible support of interoperability, bridge, Mass.: MIT Press.
even when the applications involved do not Devlin, K. 1991. Logic and Information. Cambridge,
have explicit ontologies. U.K.: Cambridge University Press.
McCarthy, J., and Hayes, P. 1969. Some Philosophical
References Problems from the Standpoint of Artificial Intelli-
Barwise, J., and Etchemendy, J. 1987. The Liar: An Es- gence. In Machine Intelligence 4, eds. B. Meltzer and
say on Truth and Circularity. New York: Oxford Uni- D. Michie, 463–502. Edinburgh, U.K.: Edinburgh
versity Press. University Press.
FALL 2003 73
Articles
AI Web Site Could berta. His current research focuses on the design and
formal characterization of ontologies and their appli-
cation to problems in manufacturing and enterprise
Use Your Help! engineering. Gruninger is also the project leader for
ISO 18629 (PROCESS SPECIFICATION LANGUAGE) within the
International Standards Organization.
74 AI MAGAZINE