Integrating multiple paradigms within the blackboard framework

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/3187712 Integrating multiple paradigms within the blackboard framework Article in IEEE Transactions on Software Engineering · April 1995 DOI: 10.1109/32.372151 · Source: IEEE Xplore CITATIONS READS 48 23 2 authors: Sanja Vraneš Mladen Stanojević 97 PUBLICATIONS 366 CITATIONS 62 PUBLICATIONS 231 CITATIONS Mihajlo Pupin Institute SEE PROFILE Mihajlo Pupin Institute SEE PROFILE All content following this page was uploaded by Mladen Stanojević on 10 January 2014. The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document and are linked to publications on ResearchGate, letting you access and read them immediately. 244 IEEE TRANSACTIONS ON SOFTWARE ENGINEERING, VOL. 21, NO. 3. MARCH 1995 Integrating Multiple Paradigms within the Blackboard Framework Sanja VraneS, Member, IEEE, and Mladen StanojeviC Abstract- While early knowledge-based systems suffered the frequent criticism of having little relevance to the real world, an increasing number of current applications deal with complex, real-world problems. Due to the complexity of real-world situations, no one general software technique can produce adequate results in different problem domains, and artificial intelligence usually needs to he integrated with conventional paradigms for efficient solutions. The complexity and diversity of real-world applications have also forced the researchers in the AI field to focus more on the integration of diverse knowledge representation and reasoning techniques for solving challenging, real-world problems. Our development environment, BEST (Blackboardbased Expert Systems Toolkit), is aimed to provide the ability to produce large-scale, evolvable, heterogeneous intelligent systems. BEST incorporates the best of multiple programming paradigms in order to avoid restricting users to a single way of expressing either knowledge or data. It combines rule-based programming, object-oriented programming, logic programming, procedural programming and blackboard modelling in a single architecture for knowledge engineering, so that the user can tailor a style of programming to his application, using any or arbitrary combinations of methods to provide a complete solution. The deep integration of all these techniques yields a toolkit more effective even for a specific single application than any technique in isolation or collections of multiple techniques less fully integrated. Within the basic, knowledge-based programming paradigm, BEST offers a multiparadigm language for representing complex knowledge, including incomplete and uncertain knowledge. Its problem solving facilities include truth maintenance, inheritance over arbitrary relations, temporal and hypothetical reasoning, opportunistic control, automatic partitioning and scheduling, and both blackboard and distributed problem-solving paradigms. Index Terms- Artificial intelligence, blackboard model, expert systems, logic programming, model-based reasoning, multiparadigm approach, object-oriented programming, proceduralprogramming, rule-based programming I. INTRODUCTION E XPERIENCE from building knowledge-based systems for real-world applications has shown that their power is most apparent when the problem considered is sufficiently complex. As the complexity of knowledge-based systems approaches the complexity of the real world, modularization becomes imperative. Due to its inherent modularity, the black- Manuscript received October, 1992; revised May, 1993. Recommended by D. Wile. The authors are with the Mihajlo Pupin Institute, Computer Systems Department, Belgrade 11060, Yugoslavia. IEEE Log Number 9407079. 009%5589/95$04.00 board architecture appears to have the greatest potential for knowledge-based system modularization. The modularity of the knowledge sources makes them ideal candidates even for distribution among multiple processors. But, modularization alone is not the answer. It makes a system manageable. To make it programmable, a proper paradigm has to be chosen. A programming paradigm can be thought of as a basis for a class of programming languages, as an underlying computational model, as a primitive set of execution facilities, or as a powerful way of thinking about a computer system. Although each among the best known programming paradigms (procedural, logic, object-oriented, rule-based) offers some advantages in comparison with the others, the corresponding disadvantage is that any of them is too narrowly focused to describe all aspects of a large, complex system. Therefore, we adopted a multiparadigm approach to complex system programming. A. Utility of Multiparadigm Systems The purpose of multiparadigm programming is to let us build a system using as many paradigms as we need, each paradigm handling those aspects of the system for which it is best suited. A special part of the Fifth International Symposium on AI held in Cancun, Mexico, was Marvin Minsky’s keynote address on “The Near Future of AI”, that set the tone of the conference in encouraging AI researchers to use multiparadigm approaches to solving problems. Minsky stressed that during the next decade, AI researchers need to move towards a hybrid approach of handling knowledge representation and solving problems. For example, one piece of the problem may be best solved by a rule-based expert system, another part by a decision support system, and so on. As Pamela Zave said [47], multiparadigm programming addresses one of the essential problems in software engineering-the drastic difference among aspects of a complex system, i.e. its heterogeneity. Jay Liebovitz [28] also pointed out that integration is the key trend in expert systems applications. The stand-alone expert system is becoming obsolete, as operational, real-world expert systems, need to be integrated with existing databases, information systems, spreadsheets, optimization packages, simulation software, or other software necessary for solving a specific problem. A lot of recent research has been concerned with multiparadigm languages and environments that allow the programmer to use more than one mode of thinking (paradigm) 0 1995 IEEE VRANES AND STANOJEVI6: PARADIGMS WITHIN THE BLACKBOARD FRAMEWORK for complex problems, but the approaches vary widely from project to project. Let us mention some of them here. Arctic from Carnegie-Mellon University is a functional and real-time programming language. While in conventional languages concurrency is expressed by creating multiple processes that change state in parallel, in Arctic, however, values have a time dimension. Many functions can be defined on overlapping time intervals to express concurrency. Concurrency is a natural byproduct of the functional properties of the language, rather than a language control construct. The C++ programming language (AT&T Bell Laboratories) supports data abstraction and object-oriented programming in addition to traditional programming techniques. A similar combination of imperative and object-oriented paradigms is provided by Smallword (IBM Research), while Lore (CGE Research Center France) combines object-oriented and setbased programming. Two modem paradigms-logical programming and functional programming-are also an active area of research in multiparadigm languages and environments. Their great similarities include applicative nature, reliance on recursion for program modularity, and providing execution parallelism in a natural manner. Their differences include radically different variable concepts, availability of higher order program entities and fundamental support for nondeterministic execution. In a successful combination, one might hope for the notational leverage and execution directness of functional programming, wedded with search guided through constraint accumulation. Many proposals have been offered on how to combine these two paradigms in a tasteful and powerful manner. One of the most interesting proposals came from the University of Utah 1291, [301. CaseDe from Case Western Reserve University merges imperative and specification paradigms and focuses mainly on the process of design and notations for describing a design. It deals not only with the purely functional design, but also with such design criteria as modifiability, performance, resource usage, and reusability. Tablog (IBM Research, Stanford University, Weizmann Institute Israel and SRI International) is a logic programming language that combines functional and relational programming into a unified framework. It incorporates advantages of two of the leading programming languages for symbolic manipulation-Prolog and Lisp, by including both relations and functions and adding the power of unification and binding mechanisms. An interesting extension of the Lucid language toward a high-level, real-time software-oriented tool with forinal mathematical semantics is reported in [ 141. The programming language Nial [23] supports several styles of programming including imperative, procedural, applicative, and lambda-free. Nial tools can be built to illustrate relational and object-oriented styles of programming. Researchers in the artificial intelligence field have recently been taking an increased interest in multiparadigm representation and reasoning systems. The first generation artificial intelligence systems were mostly monolithic, isolated and stand-alone, which prevented them from adequately addressing 245 the complexity, diversity and performance challenges of complex, heterogeneous, large-scale applications. However, most real world problems and situations are sufficiently complex to demand more than one reasoning technique to be used for solution, each technique attacking one characteristic of the problem domain. There is also a need to integrate AI solution techniques with conventional techniques. AI researchers have for most of the early years been guilty of ignoring a significant amount of research done in traditional fields like decision sciences and operations research. Furthermore, just as there is no one omnipotent general reasoning technique, there is little consensus on a general architecture for integrating diverse reasoning techniques. B. Characterization of Multiparadigm Systems Though a number of multiparadigm AI systems have been designed and studied, little effort has been devoted to comparing the systems or searching for common principles underlying them. A characterization of the multiparadigm AI systems along two dimensions has been suggested [17]-systems that employ multiple representations and those with multiple reasoning paradigms. We hereby introduce the third dimensionmultiple control paradigms. Along the representation dimension, a system can represent the same knowledge in different media-as in the VIVID reasoning systems, presented by Brackman and Etherington [ 131 and the multiple reasoners at a single layer of the CAKE architecture [35]+or it can have different representations for different kinds of knowledge-as in KRYPTON [2] and many sorted logic [5], [16], [46]. An interesting example of a language with multiple paradigms in the object framework is Orient/84 [40], which has been designed to describe both knowledge systems and systems of more general application. It has the metaclassclass-instance hierarchy and multiple inheritance from multiple superclasses. A knowledge object consists of a behavior part, a knowledge-base part, and a monitor part. The behavior part is a collection of procedures (or methods) that describe actions and attributes of the object in Smalltalklike syntax and semantics. The knowledge-base part is the local knowledge base of the object, containing rules and facts. Prolog-like predicate logic is employed to describe the knowledge base. The monitor part is the guardian and demon for the object and is described in a declarative manner. Along the reasoning dimension, a system can have different reasoners for the same representations-as in KRYPTON, KLTWO [41], theory resolution [38] and many sorted logic-r the same reasoner for different kinds of knowledge-as in the Patel-Schneider’s hybrid logic [33]. RAL (Rule-extended Algorithmic Language) from Production Systems Technology (developer of OPS5 and OPS83) adds rule-based and objectoriented capabilities to C programs. The Loom knowledge representation language, developed by Robert MacGregor and John Jen at the University of Southern California’s Information Sciences Institute combines object-oriented programming, rules, and logic programming in a set of tools for constructing and debugging declarative IEEE TRANSACTIONS models [31]. Loom uses a description classifier to enhance knowledge representation and to extend the class of useful inference beyond simple inheritance (found in most frame systems). It supports the description language (the frame component) and a rule language, and uses its classifier to bridge the gap between the two. The classifier gives Loom the additional deductive power to provide inference capabilities not often found in current knowledge-representation tools. A new version of Loom uses a RETE matcher for efficient classification, as does our multiparadigm language, Prolog/Rex [44] (used in BEST as a common substrate for multiple paradigm implementation). However, Prolog/Rex possesses capabilities not found in Loom like hypothetical, temporal and approximate reasoning, multiple inheritance over arbitrary relations, assumption-based truth maintenance, while our multiparadigm environment BEST (Blackboard-based Expert Systems Toolkit) [43] as a whole offers opportunistic control, automatic explanation, data-base gateway, the inclusion of operational research techniques, multicriteria optimization, distributed problem-solving framework, etc. Moreover, BEST is capable of integrating with diverse tools from universities and industry, combining knowledge-processing and conventional subsystems and customizing itself according to the domain-specific needs. In this it is similar to ABE (A Better Environment) from Cimflex 1201. Both BEST and ABE provide a very general and flexible base for integrating a variety of programming paradigms. However, BEST is a more fine-grained integration framework, where knowledge sources are single-paradigm programs and the blackboard itself is the integration platform, while ABE offers a variety of frameworks for designing applications (one of them is the blackboard framework), and even generic solutions to classes of problems (skeletal systems) at the higher level. The Loops [37] knowledge programming environment integrates function-oriented, object-oriented, rule-oriented and access-oriented programming. Another rich amalgamation of multiple programming paradigms is KEE [22], which integrates object-oriented and rule-oriented programming with a database management system. ART [21] successfully combines rules and frames (schemata) and provides the means for hypothetical and time-state reasoning, but lacks the facilities for complex knowledge base modularization, uncertainty management and explanation of its reasoning process (all provided by BEST). BEST represents one possible approach to integrating diverse knowledge representation and reasoning techniques (internal integration), and to integrating the basic knowledgebased paradigm with other programming paradigms (external integration), within the blackboard framework. It incorporates the best of traditional, procedural programming, rule-based programming, frames, logic programming, object-oriented programming and blackboard modelling in a single architecture for knowledge engineering. The user can tailor a style of programming to his application, using any or arbitrary combinations of methods to provide a complete solution. The deep integration of all these techniques yields an environment more effective even for ON SOFIWARE ENGINEERING, VOL. 21, NO. 3. MARCH 1995 a specific single application than any technique in isolation or collections of multiple techniques less fully integrated. Due to its flexibility and generality, the blackboard paradigm provides an excellent integration framework for composing the mentioned programming paradigms, without sacrificing either execution performances or validation capabilities of any of the participating paradigms. When a complex problem is carefully and appropriately decomposed into manageable subproblems, the best suited paradigm is chosen for each particular subproblem encapsulated within separate knowledge sources. Each knowledge source, which is now a single paradigm program, can be validated a great deal in isolation from the rest of the system. When the local correctness of every knowledge source is established, it is much easier to validate the whole system. The interaction between paradigms is provided through common data structures at the blackboard, accessed through the protocol defined at the conceptual level of the participating paradigms. The only possibility for one knowledge source to influence another knowledge source’s computation is through blackboard structures, and there is no way for one paradigm to disturb another paradigm’s syntax or semantics. A knowledge source that posts its partial solution to the blackboard need not even know which knowledge source uses it and when. The consumer remains unknown, which means that paradigms communicate opportunistically. Another less important possibility for some paradigm composition is a well-recognized procedure or function call. For instance, the Prolog compiler, through which the logic programming paradigm is incorporated, has built-in C, GKS, and SQL interfaces, which can provide escape to procedural languages, graphic packages or database management systems. Of course, in this kind of communication between paradigms, both involved paradigms are aware of the communication, since a message sender waits for the answer. Within the basic knowledge-based programming paradigm, our development environment offers a multiparadigm language for representing complex and heterogeneous knowledge, including incomplete and uncertain knowledge, and multiparadigm reasoning (backward- and forward-chaining and hybrid initiative) and control (hard-wired procedural, and flexible rule-based control). BEST provides techniques for hypothetical and time-state reasoning, assumption-based truth maintenance and facilities for building “explainability” into a system (saved descriptions and justifications for derived results), and a new system architecture that provides the efficiency and flexibility required for real-time applications. Moreover, the modularization provided by BEST is the key to achieving reusability and extensibility. The coherently structured chunks of knowledge encapsulated within the knowledge sources could be reused in various similar applications. The autonomy and loose cooperation between knowledge sources offer the possibility of modelling and testing integrations: multi-agents, multi-experts, and multipleprogramming-paradigms. It allows for the most natural problem/paradigm mapping and therefore for solving the problem in the manner similar to human experts. VRANES AND STANOJEVIk PARADIGMS WITHIN THE BLACKBOARD 247 FRAMEWORK To program a specific knowledge source, the user can select the following: rule-based programming with hybrid rule language and multiple reasoning paradigms (forward-chaining, backward-chaining and the combination of the two, adaptive control and search strategy, which can significantly reduce the amount of problem solving activities performed in an application, pattern matching through an indexing scheme or the RETE algorithm, and other possibilities for performance tuning, etc.). BEST’s facilities, not often found within commercial frameworks, provide the following powerful reasoning paradigms: - hypothetical reasoning, - nonmonotonic reasoning, - approximate reasoning, - reasoning under uncertainty, - truth maintenance system, using logical dependencies to maintain logical soundness, - explanatory reasoning, etc.; intelligent database and spreadsheet management; procedural programming, usually through the C language or any other conventional, imperative language, accessible through the C interface; frame language-the core of the BEST knowledge representation language (Prolog/REX) is its basic data structure, a concept which can be used to represent domain objects, situations, events and processes; logical programming through the standard Prolog environment; data-driven and access oriented programming through the use of demons and data-driven pattern matching (the RETE algorithm and indexing schemes are provided); object-oriented programming through the use of abstract data typing, inheritance, methods and message passing; model-based reasoning provided by the combination of object-oriented and access-oriented programming; distributed artificial intelligence. Before describing how different paradigms are implemented and integrated with each other (Section IV) we will pay attention to the key paradigm integration problems (Section II), and to the way our blackboard framework, used as an integration platform, solves these problems (Section III). Finally, we will use a simplified version of one of the prototype systems built by BEST to illustrate how a complex, heterogeneous problem could be solved successfully when a multiparadigm approach is adopted (Section V). l l l l l l l l l II. KEY INTEGRATIONPROBLEMS What are the main problems in the design of a multiparadigm system to solve a particular problem? One design issue is how the complex system is to be divided into single paradigm components. The most natural approach is to start with some distinction between problem aspects and then to seek the most appropriate paradigm for each. A paradigm for each aspect is chosen by looking simultaneously at several factors-expressiveness, runtime features, validation capabilities and the features that a paradigm offers for free that mirror the structure of the main portion of the problem and most easily encapsulate the problem-solving procedure. Another, more complex design issue involves the integration of the paradigms involved. The key integration problems concerned are as follows: communication and cooperation between the components of a multiparadigm system; control and synchronization of the multiplicity of components; validation of a multiparadigm system. Any multiparadigm system needs to have a way of representing the goals of the different single paradigm components and their returned results (partial results of the overall problem). Therefore, something like public language, common substrate, or teleological vocabulary, needs to be defined in which to express these goals and partial results. The control structures may be of various kinds; the way in which resources are allocated may be heuristic or algorithmic and may be based on either syntactic or semantic criteria. The communication protocol may be directed between subsystems or broadcast generally. Control may be centralized or distributed. Zave [47] identified three modes of paradigm synchronization+all, stream and event synchronization. Because each synchronization mode provides a different conceptual structure for integration, it is important to choose the synchronization mode that fits the communication needs of both programs. The first mode of paradigm synchronization is the only widely recognized conceptual structure-familiar interlanguage procedure or function call. The common assumptions shared by programs that communicate with call synchronization are the signatures of the calls. In stream synchronization, two programs communicate via messages sent on a one-way buffered communication channel (a stream), assuming that the stream producer and consumer agree upon message type and format. In the third mode, event synchronization, the common assumption shared by two programs includes a set of shared events. The programs constrain each others behavior because a shared event cannot occur unless it is simultaneously enabled in both programs. Using a blackboard framework as an integration platform, we introduce here the fourth way of paradigm synchronization-opportunistic scheduling, which will be further elaborated in Section III. The final important issue raised is how to validate a multiparadigm system. Validating a single-paradigm program in isolation from the other programs in the system often achieves significant reduction in the complexity of validation. Isolated validation is especially powerful when one program contains all the information relevant to a property of the system as a whole. The problem with isolated validation is that a single-paradigm program generally contains multiple nondeterministic choices. When the whole system executes, the other programs may make these choices in a correlated fashion, but in isolated validation that external information is lost. When correlated external decisions are needed for proper validation, then an isolated validation is useful only when such a decision can be simulated for testing purposes. If this l l l 248 IEEE TRANSACTIONS ON SOFI’WARE ENGINEERING, VOL. 21. NO. 3. MARCH 1995 simulation is too complicated it might be better to test the program in combination with other programs whose partial results are so critical to it. III. INTEGRATIONUSINGTHE BLACKBOARD FRAMEWORK A blackboard model is a particular kind of problem-solving model, i.e., a scheme for organizing reasoning steps and domain knowledge to construct a solution to a problem [ 121. The knowledge needed to solve the problem is partitioned into knowledge sources (KS’s), which are kept separate and independent. The problem-solving state data are kept in a global database, the blackboard. Although the blackboard model has been widely recognized as the most general and flexible knowledge system architecture, it is only just beginning to be appreciated outside the walls of academic research, which may be due in part to a lack of commercially available generic frameworks for blackboard systems development. Therefore, we built our own framework, called BEST (Blackboard-based Expert Systems Toolkit), with which we aimed not only to provide the infrastructure for building blackboard systems, but also to construct a platform for the integration of multiple programming paradigms. In this section we will describe the organizational principles used in our framework, and the ways it solves the main integration problems mentioned in the previous section. The architecture of the multiparadigm system developed by BEST (Fig. 1 gives a schematic of it) consists of two groups of individual computation agents, known as domain knowledge sources (DKS’s) and control knowledge sources (CKS’s), and two shared databases, known as the global domain blackboard (GDBB) and the control blackboard (CBB). A shell permits a heterogeneous knowledge sources structure, where each of them can be programmed within the most suitable paradigm. However, no matter what representation is used, to every knowledge source corresponds a Domain Knowledge Source Activation Rule (DKSAR) which has an if-then (precondition-action) format, where the if-part (Domain Knowledge Source Relevance Test-DKSRT) describes situations in which the knowledge source can contribute to the problem solving process, while the then-part (domain knowledge source invocation-DKSI) initiates the knowledge source execution. In the if-part, an event or a set of significant blackboard events in which the corresponding knowledge source is interested, are declared. An event may occur whenever a specified blackboard item (concept, fact, hypothesis, see Section III-A) is written, modified or deleted. The if-part can contain arbitrarily complex tests, for instance a boolean combination of atomic tests for the presence (absence) of certain slot values, facts, hypotheses (or their combination). In the then-part, the computation (traditional procedure, logic program, object-oriented program, a set of SQL primitives, and so on) or rule interpretation (whether a simple RBS, or a more sophisticated hypothetical or nonmonotonic reasoning system) to determine a desired effect (to either create a new object or change an old one), is initiated. The framework organization allows these various independent and diverse sources of knowledge to be specified and their Fig. 1. BEST’s architecture. interactions coordinated so that they might cooperate with one another (through the global blackboard) to effect a problem solution. Knowledge sources react opportunistically to the global blackboard changes produced by other knowledge sources executions and, as a result of its computation/cognition, produces new changes. The global domain blackboard (GDBB) is the central communication and paradigm integration medium used as a repository for a global data, partial solutions or pending knowledge sources. When one knowledge source is active, it posts on the blackboard all the information that it concludes is of importance in solving the overall problem. This information may trigger other knowledge sources. Using different relations (system defined inheritance relations and/or arbitrary user defined relations) we can divide the global blackboard into several panels with many levels of abstraction. Panels can be described as the collections of objects that share some common properties. For instance, if we are playing a war game [42], one panel may represent one side, and the other panel the other side. In this example we can also distinguish different levels of abstraction, if we are tracing the decision making process from the supreme command to the commands at the lowest level. For every command we know to which level of abstraction it belongs. The global blackboard does not have a predetermined structure. A user can easily tailor the structure of the global blackboard to perfectly fit a problem of interest. The global blackboard is implemented using a common substrate, declarative data carriers of Prolog/Rex (facts, hypotheses, concepts, relations, demons, and contexts, see III-A), and the user is free to use them in the way that is most convenient in solving a particular problem. Since each knowledge source contains many internal hypotheses and partial solutions which need not (or even should not) be visible to other knowledge sources, each knowledge source can contain its own local domain blackboard (LDBB). A local blackboard can also be partitioned into panels containing different types of data and can have three other dimensions (internal abstraction level, solution alternative and/or interval). All information that belongs to the local blackboard is local to that knowledge source, and cannot be used in other knowledge sources. We implement VRANES AND STANOJEi’k: PARADIGMS WITHIS THE BLACKBOARD FRAMEWORK the local blackboards in order to avoid undesirable interference between knowledge sources, and to make the mapping to the distributed or multiprocessor platform much easier and more beneficial. Since BEST was conceived as a potentially distributed/multiexpert development environment from the very beginning, we have gone to some lengths to avoid (or at least reduce) the blackboard information bottleneck. One way of achieving this is just by providing a local blackboard, shared by the interrelated rules or procedures within a single knowledge source. Two knowledge sources can share the same local blackboard in two cases, when the names of the domains are declared to be the same (the names of the local blackboards are the same), or when the domain name of the activated knowledge source is shared (the activated knowledge source works on the local blackboard of the previous knowledge source). Contention on the common memory containing global blackboard information is minimized, since a great percentage of computation and inference is done locally (about 90% of all blackboard referenees during the interpretations are to a local blackboard rather than to the global blackboard, so the possible source of contention in a distributed environment is significantly reduced). Knowledge sources can also manipulate the data on the global blackboard, and this is, at the same time, the only permitted way of communication between knowledge sources. Our blackboard framework for distributed problem-solving provides a tool for finding the cooperative solution of cognitive problems by a centralized and loosely coupled collection of knowledge sources located in a number of distinct processor nodes. The knowledge sources cooperate in the sense that none of them has sufficient information to solve the entire problem. By centralized we mean that both control knowledge sources and the common database (blackboard) are centralized in order to avoid a coherence problem and to simplify the overall problem-solving process supervision. Mutual sharing of information representing nonlocal aspects of the problems is performed in a highly controlled and constrained way. Loosely coupled means that individual knowledge sources spend a great percentage of their time in cognition, computation or decisionmaking rather than communication. The so called “result sharing” paradigm is inherent to blackboard architecture, where individual knowledge sources assist each other by sharing partial results, based on different expertise, different dimensions of the problem or somewhat different perspectives on the same overall problem. BEST can deal with a situation that changes through time by making chains of global blackboards snapshots and storing them in a historical blackboard (HBB), that can’ be used for long term analysis, debugging, tracing, replay, etc. The frequency of memorization may vary according to the current context and the relative importance of changes on the global blackboard. A context mechanism is used for modelling hypothetical alternatives, pursuing automatically and efficiently the analysis of branching possibilities. It does this by building a tree-like structure of related contexts. A system baseline (SB) knowledge source captures knowledge relevant for system configuratjon and initialization. 249 The global domain blackboard monitor (GDBBM), conflict resolution (CR), meta plan (MP), domain knowledge source activation (DKSA) and global domain blackboard indexes (GDBBI’s) will be described later in greater detail (see Section III-B). Let us now examine our knowledge representation language, Prolog/Rex, that serves as a common substrate for the implementation of different programming paradigms (see Section IV), while its declarative part represents a public language used for paradigm communication via the global blackboard. A. An Overview of Prolog/Rex Knowledge representation is central to knowledge-based systems design and their ability to integrate with other programming paradigms and software technologies. Therefore, the success of the system depends a great deal on the extent to which domain knowledge and data can be represented within a particular paradigm. The optimal knowledge representation scheme depends on the application. If a single isolated paradigm is adopted, it might diminish individual performance and erode customized advantages and efficiencies in a heterogeneous, complex domain. Therefore, we offer a hybrid, multiparadigm knowledge representation system, built on top of Prolog (Prolog/Rex) [44], that can accommodate natural heterogeneity in complex knowledge bases. Prolog/Rex, provides a substrate on which the user can easily implement many single-paradigm programs. The main benefits of multiparadigm representation and reasoning systems include the following: l increased expressive power because a single-paradigm knowledge representation language may not easily represent everything, l increased reasoning power, that is, the ability to make more inferences out of the same knowledge, l increased efficiency, because a specialized reasoner can make optimizations that a general purpose reasoner cannot. Therefore we developed Prolog/Rex, a hybrid language. which integrates several knowledge representation paradigms and offers remarkable reasoning and control flexibility, not found in other Prolog extensions. The declarative knowledge representation facilities of Prolog/Rex (DK, Fig. 2) are used to implement the structure of the global and local domain blackboards (GDBB and LDBBs, Fig. 1). Global blackboard constructs represent not only the problem-solving state data. They also contain a teleological vocabulary, representing the goals of the different single-paradigm knowledge sources, their triggering events and returned results. All the knowledge used in the reasoning process is organized in data abstractions called contexts. Conceprs, facts and hypotheses comprise the content of one context. Contexts are used as an element of a truth maintenance system implemented in Prolog/Rex. similar to a viewpoint mechanism in ART [22]. The conrexr mechanism provides Prolog/Rex with the ability to pursue hypothetical alternative pathways to a goal and to represent the situation that changes in time. 7 NO. Fig. 2. Prolong/REX knowledge representation MARCH 1995 facilities. The main portion of declarative knowledge we want to handle in the system is brought together in a common category called concept which is a frame-based data abstraction that provides the knowledge engineer with an easy means of describing the types of domain objects that the system must model. Information about the object described in terms of the concept is stored in the slots. A slot value can be inherited, i.e. pased down the object hierarchy. A restriction on slot (value and type) provides automatic consistency checking. Each slot declaration consists of two parts-a name, and contents that may change under the action of the rules or procedures. Concepts may belong to many classes simultaneously, thereby giving rise to multiple inheritance hierarchies. The inheritance network is declared by means of the two standard inheritance relations-is-a and instance-of. Is-a relation links one general concept to another, while instance-of represents one of something, i.e., a particular object that, of course, reflects all the properties which have been defined for their parent concept, but also has the properties that vary between individuals with the same parent. PrologRex automatically creates the inverse relations (has-a, has-instances), which is a labor-saving feature. Furthermore, the is-a relationship is transitive, which means that declaration of is-a relationships resulting from the transitive composition of others can be omitted. It is obvious that not all relations between conceprs should produce inheritance. Moreover, not all the possible inheritance links among concepts can be expressed with standard, system-defined relations is-a and instance-of, and their inverse relations has-a and has-instances. Therefore, the potential to create user-defined inheritance and other relations is provided by Prolog/Rex. Demons represent some means of noticing and acting upon changes in slot values, necessary for significant event detection. Thus, the execution of a specified function can be triggered whenever a slot value is referenced. or whenever a new value is stored in a slot. Such demons can automatically and invisibly compute values to be placed in a slor (Prolog/Rex’s intrinsic inference). Demons are also needed when values must be looked up in databases. They also participate in the expert system reasoning activities by providing “automatic” inference as part of each assertion and retrieval operation. They also provide a way of associating domain-dependent behavior with concepts. which, together with inheritance and data abstraction, qualifies concepts as an object-oriented programming facility. Facts are used to represent simple relational statements. A hypothesis is used to define the assumption that invokes the alternative pathway to the solution, i.e. that changes the problem state, through contexr generation. Procedural knowledge (PK, Fig. 2), i.e., the behavior associated with domain objects, is expressed in terms of production rules or classical procedures (C language, SQL, graphic primitives, etc.). Although Prolog itself is a rule-based system that uses stored facts and rules to deduce solutions, experience has shown that the Prolog built-in rule interpreter is not always efficient enough, especially when the knowledge base is large. Therefore, we built a new rule language and its interpreter. In our rule language we distinguish two basic types of rules-domain rules (forward-chaining rules, backwardchaining rules, constraint rules, Fig. 2), that generate hypotheses and conclusions about a problem area, and control rules (meta-rules, domain knowledge source activation rules-DKSARs, set-control rules) that permit flexible control of reasoning, adjustable to the processing context. The DKSAR type of control rule has a special role in solving the key paradigm integration problem-control and synchronization. An escape to the underlying standard Prolog system is provided from both sides of any kind of Prolog/Rex rule (using VRANES AND STANOJEVIt: PARADIGMS WITHIN THE BLACKBOARD FRAMEWORK the @ operator), which can be useful for building complex tests in the precondition part of a rule, for writing messages from the action part, for providing procedural attachment (accomplished through a built-in C interface) or for any other action that logic programming is suitable for. 251 is influenced by the current strategy (depth-first, breadth-first, hill-climbing, beam-search, A*, etc.). Initially the depth-first strategy with the use of rule priorities is set, but it can be changed later using the set-control rules. When the conflict is resolved according to the selected strategy, a rule instance is put into the rule agenda, and control is returned to the if-part. If there exists another set of data that satisfies the preconditions, B. Opportunistic Control and Synchronization the conflict resolution is performed again, otherwise control We distinguish two levels of built-in control in BEST. The is returned to the local monitor, and the next rule from the first level of control affects knowledge source scheduling and list is tried. The rule activation mechanism invokes the then-parts of activation (Fig. l), while the second level of control affects rule scheduling and interpretation. Whenever a concept, fact forward-chaining, constraint, and set-control rules according or hypothesis on the global blackboard is asserted, retracted, to the selected strategy. The order of rule firing can be changed or modified, a global domain blackboard monitor (GDBBM) is using the meta-rules. Meta-rules are sensitive to the state of the activated. It receives information about the performed opera- global blackboard, local blackboard and the rule agenda. When tion, and then tries to find a Global Domain Blackboard Index a then-part of a selected rule instance is fired, it can perform (GDBBI) that matches the given concept, fact, or hypothesis. changes on the global blackboard and local blackboard, thus The global index contains a list of all DKSAR’s whose if-parts transferring control to the global or local monitor. When they are affected. After the first DKSAR is picked from the list, the finish their tasks, control is returned to the rule activation control is transferred to its if-part. If all the preconditions in mechanism, and this control loop repeats. The loop terminates the if-part are fulfilled, the conflict resolution (CR) mechanism when the rule agenda is empty, or on the user’s request. Since the choice of a control structure strongly affects gains control. Its task is to put the DKSAR instance into the efficiency, we have found it extremely useful to offer the knowledge source agenda (KSA) using the DKSAR’s priority, and the value of a criterion (if one is defined), such as the choice of switching to the most appropriate control strategy efficiency, reliability, or a value returned by a knowledge for every single subproblem we solve. Prolog, with its “hard source or a heuristic function which combines these values. wired” backtracking strategy, lacks this kind of flexibility. The The agenda is always kept sorted in decreasing order of Prolog/Rex rule manager supports both forward and backward priorities. When the conflict is resolved, control is returned chaining (where forward-chaining is the basic strategy) while to the if-part in an attempt to satisfy the preconditions using the backward-chaining mechanism is activated whenever the other data. If it succeeds, control is transferred again to the pattern from the precondition part of the forward-chaining rule conflict resolution mechanism, otherwise control is returned cannot be matched with the fact from the knowledge base, to the global monitor and the next DKSAR from the list is but there exists a backward-chaining rule that can prove that fact. By allowing the interleaving of forward-and backwardtried. A domain knowledge source activation (DKSA) takes the chaining inference at the rule level, the facilities can cooperate, first DKSAR instance from the agenda and transfers control not compete. Moreover, no matter which control strategy has and the data collected in the DKSRT to the domain knowledge been selected, it is sometimes useful to control the selection source invocation (DKSI)-the then-part of the corresponding of the next rule, instead of following the knowledge base DKSAR. In cases when there are many DKSAR instances with order of rules. A human expert, for instance, albeit always the same highest priority, meta -rules can be used to determine having free access to his own knowledge, identifies and the sequence of their invocation. Meta-rules perform this task applies exactly those rules that are relevant to the problem using the information from the global blackboard and agenda. at hand. To imitate that focused manner of applying rules, Within the then-part, data either on the global blackboard or on i.e., to structure and order the unstructured world of rules, our the local blackboard, can be changed. If the data are changed agenda-based rule manager offers the possibility of assigning on the global blackboard, control is transferred to the global a priority to a rule in its original definition (regardless of monitor, and if they are changed on the local blackboard, whether the rules are preferred because they deal with the control is transferred to the local domain blackboard monitor most serious issue, or are the fastest to execute, or because (LDBBM), thus initiating the knowledge source execution. they were written by the most knowledgeable human, etc.). When either the global monitor or knowledge source finishes Apart from priorities, heuristics expressed in terms of metaits task, control is returned to the knowledge source activation, rules or set-control rules can be used to control the rule firing and the control loop repeats. The control loop terminates when order. Meta-rules enable the user to intervene in the inference the agenda is empty or on the user’s request. process, to define his own strategy of applying rules and When a change on the local blackboard occurs, the local to dynamically change that strategy according to the current monitor gains control. It uses the Local Domain Blackboard context. Unlike the similar approaches [lo], [19], [9], which Indexes (LDBBI’s) to find the names of affected forwardincorporate metaknowledge capabilities in the standard Horn chaining, constraint and set-control rules. The first name is clause interpreter, our meta-rule describes an action to be taken and control is transferred to the if-part of the correspond- undertaken by our own flexible rule manager whenever it ing rule. If all the preconditions are fulfilled, then the conflict focuses on a rule from a conflict set involved in the meta-rule resolution mechanism takes control. The order of rule firing precondition. 252 IEEE TRANSACTIONS Due to the flexibility of Prolog/Rex inference mechanisms, it is much faster with those problems where Prolog’s blind, depth-first search strategy does especially badly, i.e., with problems where combinatorial explosion creates a seemingly infinite number of possible answers (such as the possible configurations of a machine). C. Paradigm Cooperation and Integration ON SOFTWARE ENGINEERING, VOL. 21, NO. 3, MARCH 1995 TABLE I GDBB INTEGRATION PARADIGM RULE-BASED PARADIGMS INVOCATION by asserting facts or concepts on the LDBB PARTIAL GDBB RESULT ON using global structure in forward-chaining rule No matter which paradigm is used for knowledge source by referencing the 1 using built-m predicates CONCEPT programming, it is invoked using the DKSAR rule. The knowlinitial BASED edge source is initialized when a premise of the corresponding slot of a concept PARADIGMS I DKSAR is fulfilled. A bridge to the active body of the knowledge source, represented within the most appropriate usmg built-in predicates or using B structure in LOGIC programming paradigm, is provided through the action part of structures in then-part of then-part of DKSAR PROGRAMMING DKSAR the DKSAR. The knowledge sources communicate with each other using the global blackboard. Each knowledge source can be implemented using the paradigm that solves a given PROCEDURAL using embedded C in @ subproblem in the easiest way. Those differently implemented PROGRAMMING structure in then-part knowledge sources can be integrated by posting the results DKSAR of each knowledge source on the global blackboard in the uniform manner, using Prolog/Rex declarative data carriers INTELLIGENT (facts, hypotheses, and concepts). r- DBMS All the paradigms supported by BEST can be divided into five groups based upon their implementation (see Section IV). The first group contains paradigms implemented predominantly using concepts, relations, and demons: access-oriented, object-oriented programming, and model-based reasoning. The The IDBM paradigm is integrated in the BEST environment second group comprises paradigms implemented using rules. by making queries of a data-base in the if-part of a DKSAR, This group can be divided further into two subgroups contain- and asserting the wanted data on the global blackboard in the ing paradigms that use the context mechanism, and paradigms then-part of the DKSAR. that do not use it. The first subgroup contains the hypothetical, The logic programming paradigm in BEST is provided by nonmonotonic, and time-state reasoning, while the second using one predefined structure in the if-or then-part of the one contains paradigms based on forward- and backward- DKSAR. This structure allows calls to Prolog predicates. The chaining reasoning without the use of the context mechanism. results can be written onto the global blackboard either by The approximate-reasoning paradigm is a member of both using out arguments of the predicate call, or by using built-in subgroups. The third group represents the intelligent data- predicates in the Prolog code. The procedural-programming paradigm can be invoked base management (IDBM), the fourth-logic programming, and the fifth group-the procedural programming paradigm. In from the then-part of the corresponding DKSAR. The results this section we will describe the invocation of the knowledge can be written onto the global blackboard (in the then-part of sources implemented using different programming paradigms, DKSAR) by using the out parameters of the procedure call. Apart from paradigm integration via the global blackboard, and the way these knowledge sources assert their results on the global blackboard (Table I). BEST provides some means of direct paradigm communicaIf a knowledge source is implemented using one of the tions. Rule-based paradigms can invoke any other paradigm paradigms based on concepts, it is activated by referencing supported by BEST (Table II). The logic- and procedural(asserting, or modifying) a slot value of a certain concept in programming, as well as the IDBM paradigms can be invoked using structures in the if-or then-part of a domain rule. The the then-part of the DKSAR, i.e., by sending the initializing message. If we want to work on a local blackboard, then paradigms based on the use of concepts can be invoked by it is necessary to assert all the initial concepts on the local referencing a slot value of a concept in the if- or then-part blackboard and to reference the slot value of a certain concept. of a rule. Paradigms based on the use of concepts can also invoke any Partial results can be written onto the global blackboard using other paradigm. The rule based paradigms and the IDBM parasome built-in predicates. When a knowledge source is implemented using one of digm can be invoked using some built-in predicates. The logic the paradigms based on rules, then it is necessary to assert programming paradigm is directly accessible, because methods a new fact or to reference a slot value of a concept on a local are implemented in Prolog. C can be invoked from Prolog, blackboard (in the then-part of the corresponding DKSAR) in therefore the integration with the procedural programming order to initialize the reasoning process. The partial results can paradigm is provided. The logic programming paradigm is directly integrated with be written directly from the then-part of domain rules onto the all the other paradigms using built-in predicates, because the global blackboard. VRANES AND STANOJEVk: PARADIGMS WITHIN THE BLACKBOARD FRAMEWORK 253 We introduce multiparadigm control by providing both built-in procedural control and rule-based control, based on CALLING PARADIGM meta-rules. The first provides a mechanism for different search strategies (depth-first, breadth-first, A*, hill-climbing, beam-search, depth-first-iteratively-deepening, etc.). The rule-based control paradigm lets the user define heuristics (expressed in terms of meta-rules) to choose the rule firing and context expansion schedule. An advantage of the rule-based IULE-BASED using built-in using built-in ?AFtADIGMS predicates pdiCateS control is the ability to provide an external control of the C reasoning system, and to change that control dynamically. It A makes internal structures developed by the inference engine ZONCEFTusing using built-in L BASED concept L predicates accessible for inspection and modification by metu-rules. stNctures E PARADIGMS To enhance Prolog’s inherent pattern matching (syntactic unification) and resolution techniques, we used the simple D LOGIC using e direct indexing scheme first [36] (storing for each pattern a list stNctures P PROGRAMMING of rules in whose left sides it appears). By indexing we A buy speed at the expense of space. This has significantly R A PROCEDURAL using using using improved Prolog inferencing, especially when the forwardembedded C D PROGRAMMING embedded C embedded C chaining strategy is used. Due to its simplicity, the indexI ing scheme is efficient when the number of facts that are G to be matched with appropriate patterns is small, but it using sql using built-in M INTELLIGENT using built-in DBMS structures predicates predicates shows rather poor performance in the case when a large I I number of facts correspond to each pattern. Therefore, we implemented the heavily-modified RETE pattern matching algorithm [ 151 which shows better performance in these cases. BEST environment is implemented in Prolog and escape to This is caused by the fact that the order of complexity function for the RETE algorithm is lower than the order of the underlying Prolog is possible from any point in BEST. The procedural programming and IDBM paradigms cannot complexity function of the indexing scheme. The conclusion is obvious-when we have a large number of facts that directly invoke other paradigms. match a particular pattern, the RETE algorithm is recommended. Otherwise, it is better to use the simple indexIV. IMPLEMENTATIONOF THE SYSTEM ing scheme. This observation leads to the conclusion that, Let us now see how different programming paradigms are due to the performance tradeoff between different pattern implemented using Prolog/Rex (BEST’s knowledge represen- matching algorithms, the best solution is to offer the user tation language) as a common substrate. the possibility to choose the technique most suitable for a particular knowledge base, or even to combine dynamically A. Rule-Based Programming two different techniques. Our shell offers this kind of flexiModern knowledge-based systems use a large number of bility. Let us now describe how different, powerful reasoning rules about a problem area or domain. Even if the knowledge base is modularized around the blackboard framework, effi- paradigms are supported by BEST. ciency of the rule manager can enormously affect the success 1) Truth Maintenance: An additional important feature of of the system. The runtime performance in knowledge-based BEST is the ATMS-like [6], [7], [8] truth maintenance capasystems depends on two factors: the appropriateness of avail- bility implemented using the context mechanism (the correable control structures for the domain (that is, adaptability) and spondence between terms used in ATMS and BEST is shown the ability to run the system with minimal levels of intervening in Table III). In BEST, constraint rules are used to detect a language interpretation (that is, compilation). With respect to contradiction in a context, handled by the poison function. The the former, BEST provides many reasoning methods, control believe function contradicts all competing solutions. When a structures and rule application strategies, while with respect to new datum is asserted in a context, its label is updated, and the latter, it provides two rule expansion methods (an indexing data that can be inferred subsequently will be asserted when scheme and a heavily-modified RBTE expansion method). The the corresponding rules fire. If the asserted datum causes a RETE algorithm replaces repeated testing of condition clauses contradiction, a constraint rule will fire, thus poisoning the with the maintenance of each rule’s current applicability as current context and all of its descendants. The labels of all working memory changes. facts, hypotheses, and concepts contained in the poisoned The best contemporary environments for building contexts will be changed, and the rules whose precondition knowledge-based systems sometimes provide a rich parts were satisfied in the poisoned contexts will be removed amalgamation of knowledge representation formalisms, from the agenda. Dependency-directed backtracking is avoided and/or offer a powerful multiparadigm inference mechanism, since there is no need to search for the assumptions underlying but severely limit control specification or external control. a contradiction, then to select a culprit in order to resolve TABLE II DIRECT COMMUNICATION 254 IEEE TRANSACTIONS TABLE III ATMS-BEST CORRESPONDENCE ATMS nodes assumptions contexts labels BEST facts,concepts hypotheses contexts values of slots fact-in, hypothesis-in environment values of the slot has_hypothesis nogood database no-megegairs fact the contradiction by performing a context switch. After a contradiction is detected in BEST, and the inconsistent context poisoned, the context switch is performed simply by transferring the control to the rule that operates on some other context. Moreover, TMS-like nonmonotonic justifications [ 1l] which are not supported by ATMS, can be encoded in BEST using forward- or backward-chaining rules and constraint rules. In BEST we assert a new datum (using forward- or backwardchaining rules) only if all nodes of an inlist [ 1l] are in a current context and if the nodes of an outlist [ 1l] are out. The constraint rule is used to detect a contradiction if any of the nodes from the outlist becomes in, or vice versa. Hypothetical and nonmonotonic reasoning are facilitated using the BEST context mechanism. 2) Hypothetical Reasoning: The availability of a hypothetical reasoning paradigm for commercial application is quite recent [21], 1221, although it is very helpful when there is uncertainty in determining a solution to a problem. The BEST context mechanism (see Section 1II.A) similar to the ART viewpoint mechanism, represents a set of facts that are entailed by the assumption expressed in terms of the BEST hypothesis facility, which helps in differentiating between actual facts and those that have only been presumed to be true. To reduce the number of contexts created, BEST has a facility to poison the contexts within which some predefined constraints (expressed in terms of constraint rules) are violated, and prevent the exploration of forbidden contexts. A context can generate descendant contexts that inherit all the facts true in the parent context. Offspring may selectively delete the inherited facts if desired. To keep the context structure as simple as possible, BEST has the possibility of automatically merging two or more existing contexts. 3) Nonmonotonic Reasoning: The same context mechanism (described above) is used for representing the situation that changes over time, i.e. to provide time-state nonmonotonic reasoning. Unlike the monotonic hypothetical reasoning, where the number of facts is constantly growing, in time-state reasoning the facts are retracted frequently, i.e. the number of facts pulsates, which makes this kind of reasoning nonmonotonic. ON SOFIXARE ENGINEERING, VOL. 21, NO. 3. MARCH 1995 BEST can construct a chain of contexts, reflecting the state space that changes through time. 4) Approximate Reasoning: The approximate or satisficing approach can be viewed as an attempt to construct a problemsolving system that produces the best possible answer in a given amount of time. An answer may be more or less certain, precise or imprecise, and complete or partial. This method allows the system to trade-off solution completeness, precision or certainty to satisfy time constraints. Thus, by ignoring some aspects of the solutions, or by not determining exactly some solution parameters, or by not considering all supporting or refuting evidence, these methods can produce the answers at a given deadline, but might not have an answer if interrupted before the deadline. BEST, by contrast, provides the ability to commit to an action almost instantaneously, but allows the quality of that decision to improve as long as time is available. Once the deadline is reached, the best decision arrived at is provided. BEST now supports a great variety of search strategies and their combinations to be employed in the reasoning process [39]. By using these strategies the process of finding the satisfactorily “good’ solution could be implemented simply by employing the heuristic functions that guide the search process. These adjustments can be implemented so as to enable the enduser of the specific application to reach the requested state as quickly as required. BEST offers the following methods of relaxing the requirement for the optimal solution: Adjusting the weights of g and h functions, as suggested by [34], we get the following: dynamic weighting [34]-this algorithm finds the solution whose cost is not greater than (1 + E)C*, where E is the given satisficing threshold, and C* is the cost of finding the optimal solution; AZ-an algorithm using search effort estimates. This algorithm is E-admissible, i.e. it always finds the solution whose cost does not exceed (1 + E)C* ([34], Theorem 13, p. 89); R,*-a limited risk algorithm that uses the information about the uncertainty of h. Another approach in solving the satisficing problem in knowledge based systems which is also implemented in BEST, with respect to its specific properties, is given in [32]. As stated there, “the purpose of satisficing is to resolve as few initial condition and goal states as can be managed consistent with reaching a satisfactory answer”. This system fulfills absolutely the requirements given above, and, by applying the most important rule at each stage in the decision process, minimizes the effort for finding the most satisfactory answer. To fulfill the mentioned requirements, the following steps are performed: l l l l l l l a certainty value is given to the rule premise; each rule has its evidence value, i.e. the attenuation value between 0 and 1; for each rule the satisficing thresholds are defined: this is the evidence value over which the rule is considered to be true; for the sake of simplicity it is required that each rule have only or, and, or not logical couplings between premises. VRANES AND STANOJEVIC: PARADIGMS WITHIN THE BLACKBOARD FRAMEWORK The basic heuristic is to try to resolve the most promising state first, i.e., the state that will most likely have the largest final value. The values are computed bottom-up (from the initial set of facts to the conclusions upwards) by applying uncertainty management (certainty factors, fuzzy logic, probabilistic approach) to the values listed above. Uncertainty management using certainty factors in BEST is similar to the management of certainty factors in MYCIN [4]. To every piece of evidence in the premise of a rule, i.e., hypotheses,fuct or slot value, a certainty factor can be attached (a value between -1 and 1). In the rule premise, evidence can be and-ed, or-ed, or negated. A rule itself can have a certainty factor that describes the relationship between the premise and the conclusions of the rule (cf= -1 means that if the rule premise is true, then the conclusions are false and vice versa). The certainty factors of rule conclusions are evaluated using the resulting certainty factor of the rule premise and the certainty factor of the rule. Fuzzy reasoning [25] is a well-defined reasoning system based on fuzzy set theory and the theory of fuzzy relations. We have implemented different formulae for the fuzzy set complement of the Yager and Sugeno class, for set intersection, and for union of the Schweizer & Sklar, Hamacher, Frank, Yager, Dubois & Prade, and Dombi classes [25]. We use the set complement to evaluate the membership grade for the negated evidence, set union for the or combination of evidence, and set intersection for the and combination of evidence. We use fuzzy relations to describe the relationships between the evidence in the precondition part of the rule, and the conclusion in the action part of the same rule. It is necessary to define one rule for each row of the two-dimensional array. BEST also allows the use of fuzzy functions. The incorporation of fuzzy logic qualifies BEST for the design of rule-based fuzzy industrial controllers. The degree of control achieved using the fuzzy rule-based approach could be superior to a strictly analytical approach, where there is a lot of imprecision in the controlled system no matter whether the reason is a degraded sensor, nonlinearity in the system dynamics, or delays in control signals. BEST provides the use of probabilistic theory in uncertainty management [31]. To every piece of evidence in the rule premise a probability value can be attached. Those values can be combined using the independence assumption, conservative, or liberal approach. The probability value attached to a rule describes the probability that conclusions are valid if the premise is valid. The resulting probability is evaluated using the Bayesian formula, the probability of the rule premise and the probability of the rule. Both fuzzy logic and probabilistic reasoning are used for the implementation of the BEST-based investment advisory expert system (INVEX), described in Section V. 5) Temporal Reasoning: Our development environment has several different built-in generic facilities for temporal information representation and manipulation, that can be adapted to the needs of individual application. Information about domain objects is kept in the concept database, where attributes of the object, represented by concept’s slots can be given a time dimension, i.e., can be represented as a function of time. 255 An absolute time line is implemented, which is the same for all attributes that have a time dimension. The functions that describe a character of attribute time dependency are implemented in terms of demons attached to the slot. The demon itself can be realized as a procedure in C or in any other imperative language, or as a set of rules defining the attribute value’s validity in time, or a character of time dependency. When the slot with a time tag is accessed, the demon is activated which checks the current time against the value of the time validity descriptor, and retrieves the attribute value only if it is valid at that particular moment, or computes the attribute value if it is defined as a function of time of any kind. Moreover, if the value is defined as a discrete function of time, several ways of covering unspecified time values could be defined (the value is null at time not specified or the value is interpolated if it falls between two time tags, or the value corresponding to the preceding time tag is retrieved, etc.). The values of time-varying attributes are computed on a “whenneeded” basis, which saves time and space significantly, in comparison with computing and memorizing the values for each time step. Often, a large amount of information cannot be linked to a precise time, but the duration of data or facts is well known, or these facts can be related to one another. Using BEST we can handle these situations in two different ways. One possibility is to define the time boundaries symbolically (a start-date, an end-date, a duration, a symbolic constraint) and to maintain the symbolic time line which defines the order of symbolic points. While slot inheritance is provided along the concept taxonomy in BEST, time constraint propagation can be obtained automatically. Another way of establishing the time relationships between the objects is by “user defined” relations. An arbitrary number of time relations, such as those introduced by Ladkin [27] can be defined: l always, sometimes, l equals, disjoint from, l precedes, follows, l meets, met by, l overlaps, overlapped by, l contains, contained by, l begins, ends, 9 begins with, begins before, begins after, ends with, ends before, ends after, etc. When a new interval relation is entered, all consequences are computed. This is done by computing the transitive closure of the temporal relations as in [l]. B. Data-Driven Processing Paradigm Through the RETE pattern matching algorithm, a datadriven processing paradigm is introduced in BEST. The RETE match algorithm [15] is a method for comparing a collection of patterns to a collection of objects in order to determine all possible matches, without iterating over the patterns. The iteration is avoided using a tree-structured sorting network for the patterns. Data-driven processing has become the standard in highperformance reasoning architectures and inference engines. 256 IEEE TRANSACTIONS NASA has benchmarked many inference engines and determined that data-driven are dramatically more efficient than procedure flow inference engines. Empirical evidence concerning commercially deployed expert systems also indicates that data-driven architectures are better choices for solving real-world problems. To implement the RETE algorithm we build two networks-a pattern- and join-network, for each forward-chaining and backward-chaining rule. In the pattern-network we try to match patterns with facts, and in the join-network we examine whether all preconditions of one forward-chaining, or backward-chaining rule, are satisfied or not. Our RETE algorithm is much more complex than the original one [ 151. The original algorithm was developed for use in OPS5, a first generation expert system language which supports only object-attribute pairs, forward-chaining rules for knowledge representation, and the best-first search strategy. Providing a variety of formalisms to represent domain objects over which a Prolog/Rex rule operates, and a variety of inference paradigms (forward- and backward-chaining, hypothetical and time-state reasoning, monotonic and nonmonotonic reasoning, negations, different search strategies, uncertainty management, etc.) required a significant extension to the original RETE algorithm. Similar pattern- and join-networks are built in the LISP implementation of the RETE algorithm in ART [2 11. Although we do not know ART in detail, it seems that implementations differ a lot in network topology, the process of variable bindings, the partial matches manipulation in hypothetical reasoning, etc. An additional major difference induced by certainty management in BEST is its inability to build separate pattern-networks for or-ed patterns. This splitting rather simplifies the network topology in ART. While a rule precondition part is represented in terms of its corresponding pattern- and join-network, an action part is realized in exactly the same way as in the indexing scheme. C. Object-Oriented Paradigm Combining object processing with knowledge-based technology is a recent and powerful innovation in data processing [26]. Individually, each of these techniques has many advantages over conventional programming technology, and when used together the advantages are enhanced even further. Rule processing provides support for representing expertise, inference, hypothetical reasoning, and truth maintenance, while object-oriented processing provides support for data abstraction, knowledge encapsulation, inheritance, reusability, and extensibility. The message passing capability of objectoriented processing allows the system to keep knowledge about data separate from knowledge about reasoning, which is critical for good data abstraction and the encapsulation of knowledge, while pattern-matching rules and data-driven reasoning, accomplished through the RETE algorithm, enable a clear and concise specification of the algorithm. In “object-oriented” programming the central entities are data, not procedures or rules. The data objects invoke one another through messages to each other. Collections of objects n ON SOFIWARJZ ENGINEERING, VOL. 21, NO. 3, MARCH 1995 that share some properties are implemented through classes. A class describes the methods that capture behavior of its instances. Classes implement a very fundamental concept in object-orientation, namely, abstract data typing. The second powerful object-oriented concept is inheritance. Inheriting behavior enables code sharing and reusability, while inheriting representation enables structure sharing among objects. The combination of these two types of inheritance provides a powerful modelling and software development strategy. The third powerful object-oriented concept is object identity, i.e., the property of an object that distinguishes one object from another. I) Data Abstraction: An object in BEST, expressed as a concept instance, is a self-contained entity consisting of its own private data contained in slots, and a set of operations, implemented through demons. These operations constitute a protocol that provides the object’s external interface to the rest of the system. The algorithm used to implement each of the demons attached to the particular concept’s slots is encapsulated within the concept, and is hidden from the users. All interaction with an object occurs through messages sent to it requesting an operation in its protocol. A message is sent when a slot (to which a demon that realizes the method is attached) is named. Methods are implemented through the procedural parts of demons, while the declarative parts of demons are used in the process of dynamic binding. A method has the exclusive capability to manipulate a particular object’s private property contained in the slot (performing the computation and returning a value, sending the message, and so on). The exact mechanism used by the object to respond to the messages, implemented in demon is encapsulated within and is not visible outside. Users need know only how to send the appropriate message, which is rather simple in BEST (slot naming). Classes, as well as objects, in BEST are represented by concepts. Every concept has a unique name and an arbitrary number of slots. Slots in concepts representing classes describe their structure, and play the role of instance variables. The slot values can be the default slot values for the objects-members of the corresponding class. Strong typing can be provided using demons that will examine the slot value whenever it is instantiated, or modified. The new concepts in BEST can be created either statically during concept taxonomy creation and initialization, or dynamically, using rules or demons. A new concept can be created using the addc predicate with the name of this concept as the only argument. If this new concept represents an object or a class, a slot defining a relation with the inheritance property must be added (the insert predicate, described later, performs this task). Using the inheritance mechanism, the new concept inherits the prescribed structure. The predicate remc is used to delete a concept with a given name, and the predicate changec changes an old name of a concept to a new one. BEST explicitly facilitates class extensions. To define inheritance between concepts, a user must define corresponding relationships between them by adding slots that represent relations with the inheritance property. Every subclass, or every member of a class, is mentioned as the slot value (representing a particular relation) of the corresponding class. VRANES AND STANOJEVIt: PARADIGMS WITHIN THE BLACKBOARD FRAMEWORK Methods (procedural parts of demons) in BEST are implemented as Prolog clauses. The selector, or name of the method, represents the functor of the head, the first argument is the name of the class or object-the owner of the method, the second argument represents the name of the target object, and the third argument of the head is a list that contains values (as formal parameters). The body of the clause performs the desired operation of the method. Methods can be invoked implicitly, by accessing a slot value of a concept, using a mechanism for dynamic binding, or explicitly, by calling the corresponding method. There are four basic operations that can be performed on concepts: - get a slot value, - insert a new slot value, - modify a slot value, and, - delete a slot value. To each of those basic operations corresponds one predicate in Prolog (hgetconc, hinsert, hmodify, and hdelete). These predicates check first if there exists a demon that performs the requested operation (using the message predicate) and invoke it in the case of existence; otherwise the default operation is performed (defined by predicates getconc, insert, modify, and delete). Each of these predicates has three arguments where the first argument represents a name of a concept, the second one-a slot name, and the third one-a list of slot values. The predicate message is used to implement the dynamic binding of message selectors to the methods that implement them. The message predicate has four arguments. The first argument represents a name of a target concept, the second one stands for a slot name (to which a demon is attached), the third argument is a list that contains the slot values, and the last one describes the desired operation (get, insert, modzfy, delete). This predicate checks first to see if there is a demon attached to the slot of the target concept. The demon can belong to the target concept, or it can be defined in a class concept that is linked with the target concept by one of the relations with the inheritance property. This relation must be mentioned as the value of the slot relation in the declarative part of this demon, whereas class concept must be a value of the slot concept. The operation specified in the call of the message predicate must correspond to the value of the demon’s slot operation. If all these preconditions are met, then the clause that implements the procedural part of the demon is called. This call is created using the functor which is the value of the slot method, and the three arguments (the name of an owner concept, the name of a target concept, and the list of the values). The third way to invoke a method is to explicitly call the predicate that implements it. A user must specify the functor, and the three arguments of the call that match the head of the predicate. The names of demons can be overloaded, which means that different concepts can invoke the demon using the same name, but with different semantics and implementations. The demon name can be overloaded even within the same concept, in the sense that for every slot operation there can exist a different 251 method. Method names can also be overloaded; i.e. different owner concepts can share the same method name. Demons can be used to implement constraints that test the correctness or completeness of the abstract data types represented by concepts. They can test the slot values, or have some pre- and postconditions that must be satisfied before a particular method is executed. 2) Inheritance: Another property that makes a language “object-oriented” is inheritance. In BEST, objects (implemented in terms of concepts) may be defined as members of one or more classes. The class defines the internal form of every object of the class and the methods associated with objects of the class. New classes can inherit both representation (attributes, instance variables, and so on) and behavior (demons that realize operations, methods, messages, and so on). Inheriting representation enables structure sharing among data objects, while inheriting behavior enables code sharing. Inheritance also provides a very natural mechanism for “taxonomizing” concepts into well-defined inheritance hierarchies, which is especially suitable for blackboard object organization. Moreover, inheritance introduces a semantic network representation paradigm in BEST. In the semantic network representation, nodes represent concepts and links represent inheritance relationships or any other user-defined relationship which might exist among objects (includes, works in, supervises, consists of, and so on). Concept may belong to more than one class, i.e. multiple inheritance is provided. Using inheritance in concept definitions assists in defining object protocol in a relatively standardized way. The inheritance provides a facility of polymorphism, which enables uniform treatment of objects from different classes. Using classes and inheritance provides a simple and expressive model for the relationship of various parts of system definition and assists in making components reusable or extensible in system construction. The getconc predicate, that searches the slot values of a concept, checks first whether the slot name is one of the defined relations, then tries to find the slot in the given concept, and if that fails, it looks for the slot in the ancestor concepts. This implementation has the following consequences: - slots that represent relations cannot be inherited; - the slot values of the ancestor concepts are overridden, if the slot is defined within the given concept; - if on the path from the given concept to the root of the hierarchy tree there exist several definitions of the same slot, then the definition that is the closest to the given concept is taken. It is interesting to note that BEST provides flexibility in overriding the type of a slot value. In BEST, demons are used to provide strong typing. If, for one slot, the same demon is used for all descendant concepts, then no redefining of the slot value type is allowed. The arbitrary redefining of type is also provided by definition of overriding demons attached to the same slot in the descendant concepts. The inheritance property of a relation can be limited to allow the inheritance of a predefined set of slots. If we omit a slot in this set, then this slot will not be inherited. In the same hierarchy tree we 258 IEEE TRANSACTIONS can have different definitions of the same slot with different demons attached to each of them, thus providing the hidden definitions. The specialization in the hierarchy tree can be expressed by defining the new slots specific for the given concept. If so specified, these new slots can be inherited further in the hierarchy tree. The demons and the corresponding methods can also be inherited. The demon inheritance utilizes the mechanism of dynamic binding. The declarative part of a demon contains the slot relation with a value that represents a name of a relation with the inheritance property, and the slot concept that contains the names of concepts in the hierarchy tree ordered by the given relation. In order to apply a method on a slot value of a given concept, the values of the slot concept are searched in an attempt to find the name of the given concept or, if that fails, to find a concept which is an ancestor of the given concept. If the demon can be inherited from any ancestor concept to any descendant concept, then the corresponding relation must be declared transitive. In order to provide demon overriding, a rule is defined and must be followed when defining the sequence of the slot concept values. The concepts that lie on the level closest to the root of a hierarchy tree must be defined first, then the concepts on the next level and so on. These values are searched in the inverse order, and the rule allows us to find the method belonging to the closest ancestor of the given concept. If the demon is defined for the given concept then it overrides all inherited demons, otherwise, the demon of the closest ancestor is inherited. An inherited demon can be excluded using the same mechanism of demon overriding. In fact, concepts, relations, and demons comprise a prototype system, because the concepts can be individually created, no matter whether they represent a class or an object. The relations provide a means for a more flexible form of inheritance (as defined in object-oriented programming paradigm) called delegation, where arbitrary concepts can inherit from one another. The inheritance relationship can be established dynamically, while the inheritance relationship of class-based languages is established and fixed when a class is created. Multiple inheritance is another feature provided by BEST. The possible multiple inheritance conflicts are resolved using a linearization strategy, or qualified slots and methods [24]. Using the getconc predicate, an inherited slot value can be obtained. The getconc predicate is backtrackable, and so all multiple inherited slot values can be retrieved. This sequence is defined implicitly by the order of relation definitions. The slot value defined in the concept in a hierarchy tree determined by a relation that is most recently defined will be found first. The same predicate can be used to directly access the slot value in the concept where it is actually defined. The third way to resolve the multiple inheritance conflict is to attach a demon (that explicitly defines a resolution strategy) to the slot. The linearization strategy for the multiple inherited demons is implemented in the process of dynamic binding. The most recently defined demon will be activated first. Another approach is to explicitly call the specific method with the known owner concept. ON SOFDVARE ENGINEERING, VOL. 21, NO. 3, MARCH 1995 3) Object Identity: In BEST object identity is assured through the unique concept names. Each concept is represented by a separate hush table. There are three equality predicates that can be used in the object-oriented programming paradigm: identity predicate, shallow equality predicate, and deep equality predicate. The identity predicate is Prolog’s built-in predicate, while the other two can be implemented using the getconc predicate. Shallow and deep copy are not explicitly provided, but can be programmed using the getconc, addc, and insert predicates. 4) Integration of Rule-Based and Object-Oriented ProgrammingParadigms: The rule-based and object-oriented programming paradigms are fully integrated. That means that slot values of a concept can be retrieved, inserted, deleted, or modified by rules (concept, assert, retract, and modify structures, respectively), and that rules can be invoked from any method. When concepts are used in rules, then the multiple inheritance problem can arise. One solution to this problem is qualified slots. The slot name can be qualified by the name of the concept which is the owner of the slot value or demon (owner-conceptname ownsslot slotname or ownerxonceptname owns-demon slotname). If the slot name is not qualified, then the linearization strategy is applied. For the multiply-inherited slot values, a demon can be defined to resolve the conflict. There are some interface predicates that allow methods to invoke rules. The getgoal predicate invokes backwardchaining rules. This predicate has three arguments, the first one stands for the first term of a goal, the second one for the second term, and the third one is a list that contains other terms in the inverse order. The predicates cassert, cdelete, and cmodify perform the specified operation on a slot value (assertion, deletion, modification), and activate the pattern matching mechanism. The precondition parts of DKSAR, forward-chaining, constraint, or control rules can be satisfied, and later fired performing the desired action. D. Access-Oriented Programming Paradigm The access-oriented programming paradigm is fully integrated in BEST. An access-oriented paradigm is based on annotated values that associate annotations with data. Fetching or storing data can cause a procedure to be invoked. It has historical roots in languages like Simula and Interlisp-D, which provide ways of converting access to a record into computations for all records of a given type. More immediate predecessors are the ideas of procedural attachment from frame languages like KRL or KL-One. Unlike the object-oriented paradigm where the object that receives a message changes its data as a side effect, in access-oriented programming, when one object changes its data, a message may be sent as a side effect [37]. BEST demons attached to concept slots are the basic computational mechanisms of access-oriented programming. Demons are invisible to programs that are not looking for them. Moreover, since a demon itself is a framelike construct, which has its own slots for saving the state and can have a demon attached to its slot, the demons are recursive. Demons can be organized in an inheritance lattice, VRANES AND STANOJEVIC: PARADIGMS WITHIN THE BLACKBOARD FRAMEWORK due to the possibility of attaching inheritance relations and user-defined relations to their slots. Demons have very low computational overhead, and can be used for both property annotations and active values. Property annotations provide a way of attaching extra descriptions to data for guiding its interpretation, for instance, for checking data types and constraints. Before storing new data into the slot, they are checked against the constraints (e.g., data type) through a demon attached to the slot. Property annotations can be nested, supporting the notion of descriptions of descriptions. Active values convert a variable reference to a method invocation. When a slot is accessed, computation is triggered, which eliminates the need for conventional functional interfaces for changing the monitored slot. Like properties, active values can be nested (which comes for free due to the frame structure of the demon), providing an automatic, invisible propagation of side effects, which makes a solid basis for model-based reasoning. Obviously, all basic concepts of access-oriented programming can be found in BEST. The same features used to enable object- and access-oriented programming paradigms, provide the model-based reasoning paradigm [ lg] in BEST. V. AN ILLUSTRATIVE EXAMPLE We have used BEST to build two prototype multiparadigm applications: the tactical decision making system for Air Forces [42], and the INVEX-Investment Advisory Expert System [45]. To build our tactical decision making aid we used knowledge-based programming as the basic paradigm, while appreciating some old methodological friends, like linear programming which was used for programming the knowledge sources dealing with route planning and different tactical manoeuvers estimation. Similarly, purely numeric methods are used in a knowledge source that calculates damage expectancy against a target, etc. This system is admittedly somewhat too complex to be used for a case study, and therefore we will take a heavily simplified version of INVEX in order to describe how a multiparadigm system could be constructed using BEST as a basis. The user sees INVEX as a monolithic data processing program because the only connection with it is the well known Microsoft Excel spreadsheet. He is not aware of the embedded rule-based system, fuzzy logic, or multicriteria analysis (an operational research technique). BEST glues all these paradigms together (spreadsheet, rule-based programming, fuzzy-logic, probabilistic reasoning, operational research) and takes care of their communication. To establish local correctness, functional testing is applied to every knowledge source in isolation. During the testing of the integrated system, only minor correction and tuning of the participating single-paradigm knowledge sources have been done. The primary motivation for the development of the BESTbased investment advisory expert system (INVEX) was the lack of a satisfactory aid to guide the project analyst and investment decisionmaker in their choice among alternative projects and project designs. The approach adopted here is 259 to decompose the decision process into stages (strategies), specify a particular methodology for each stage, and choose the most appropriate paradigm for programming the corresponding knowledge source. For the purpose of project analysis, we have adopted a stage-by-stage integrated graphical approach which uses standard financial tables based on an integrated documentation system. This stage is fully implemented using the Microsoft Excel spreadsheet. Moreover, we use the Excel spreadsheet as a front-end interface, being aware of the fact that the spreadsheet metaphor is a very good automation of what people would usually do with business data, paper, pencil and calculator, and that it has a well-known and wellaccepted form. Both BEST and Excel run under the Microsoft Windows 3.1. environment, with intrinsic gateways among the applications (using DDE-Dynamic Data Exchange protocol). The summary tables for the alternative projects and project designs are communicated from Excel (through DDE) to the global blackboard of the BEST-based intelligent server in the background that provides value beyond automation. It uses human expert knowledge for heuristic investment classification and ranking, blended with conventional multicriteria analysis and risk assessment methods. Different methodologies are used for different analyses and different programming paradigms are respected for the corresponding knowledge sources implementation. Intermediate data and partial results and decisions are communicated among knowledge sources through a shared data repository built out of the common substrate (Prolog/Rex declarative data carriers) and located on the global blackboard. Standard financial tables, combined with graphical analysis supported by the Excel spreadsheet, lead the analyst in logical stages from a standard financial analysis to a complete economic evaluation of a project and its quantifiable impacts. Various static and dynamic measures of the project effectiveness are computed in order to shed light on the project’s desirability from a different angle. No single measure can by itself provide sufficient information for judging the merit of a project, for each measures this merit from a different point of view. The values of all these measures are communicated to the knowledge source that performs multicriteria analysis (through the project summary table on the global blackboard loaded from Excel through the DDE interface), so that all the impacts can be summarized by using some weighting scheme. Of course, since there is no universal agreement on the weights assigned to the different measures of the project desirability; a dialog between the project analyst (and/or decisionmaker) and INVEX is needed in order to establish the weights that properly reflect the decisionmaker’s wishes and intentions. During a consultation, INVEX first asks about a customer’s wishes and requirements, then builds up a customer profile, where the information asked from the customer depends heavily on his intentions and the course of the consultation. These wishes and intentions are translated using production rules (another rule-based knowledge source) into the weights assigned to the different objectives in the multicriteria analysis knowledge source. INVEX is fed with data through the interface that most users already know-the spreadsheet. The Microsoft Excel spreadsheet plays the role of a user-friendly, well-known 260 IEEE TRANSACTIONS and well-accepted front-end for data entry, standard financial table generator and translator, and as a client of the BESTbased intelligent server, performing background intelligent decisionmaking activities. This client-server structure seems to be the solution for future decision support systems. Excel is also used for the presentation of the results, since it is highly graphic, with very good presentation capabilities. As a client, BEST is fully responsible for conversation flow and synchronization. This task can be successfully fulfilled through the rules only. The expert system shell is therefore expanded with a set of additional predicates to support DDE. Data being taken by BEST appear as Prolog variables instantiated after the call to a DDE communication predicate. Similarly, data being sent from BEST to a server appear as input parameters of the corresponding communication predicates and must be instantiated before the call. The rule that makes use of these predicates converts gathered data to a suitable knowledge representation form, i.e. writes them into the slots of a concept or asserts them as simple facts on the global blackboard. As an intelligent server, BEST makes some of the knowledge base elements visible to other applications. Prolog/Rex declarative data carriers can be easily interpreted by other applications: concepts and slots have close analogies in records and fields of an ordinary database. Therefore, we decided to make some of the concepts visible to other applications. This visibility is fully transparent for the inference engine. Transparency is achieved in two ways: organization of the shell guarantees that an explicit query from the client will be serviced instantly, no matter what the current activity within the shell. Second, every slot taking part in a DDE conversation has an associated demon. The demon is activated whenever a slot value changes. Activation of the demon invokes a communication predicate which in turn notifies client applications (linked to BEST with “warm” links) that a slot value has been changed. In our multiparadigm decisionmaking system we rely on expert economics judgmental heuristics for a simple test for the financial acceptability of a project and classification and ranking of alternative projects and project designs. Using four dynamic measures of project desirability, all projects are divided into five groups. Investments from the group VERY GOOD are accepted for the multicriteria decisionmaking (MCDM), while investments from the group VERY BAD are rejected. For the investments from the groups GOOD, MEDIUM, and BAD, a sensitivity analysis is performed, and then the user is asked whether to accept or reject each of these investments. This classification reduces the number of projects that will take part in MCDM by rejecting the bad choices. If specified, a risk analysis is performed on the accepted investments, and then the MCDM gives the optimal combination of investments for the given resources. Total preorder, used in MCDM, suggests the best investments from the set of accepted investments. Since the dynamic criteria used for classification are not precisely determined, we rely on fuzzy-set theory to describe the criteria and classify the projects. Each criterion is represented by one fuzzy set, as well as each group. An investment Relative ON SOFTWARE ENGINEERING, Net Present Value of Investment Return 1995 on Investment Critical Point Payback Period Fig. 3. VOL. 21, NO. 3, MARCH Membership grade functions. can belong to many groups simultaneously, but with different membership grade values, thus making boundaries between the groups fuzzy (see Fig. 3). In Fig. 3, ci stands for a compound interest, p0 for the referent payback period, and CO for the referent period of achieving the breakeven point. We determine the membership grade values for the five groups using the composition operation: IoR=O where I represents the input vector, R represents the fuzzyrelation, and 0 represents the output vector. The composition operation is similar to the matrix product where each production is replaced by the min function, and each addition by the max function. Another perspective of investment decision making relates to the issue of future uncertainty and its consequences for planning and decision making. However, high returns are often associated with high risks. An important role of INVEX is to aid managers in assessing various future alternatives and the levels of risk and return associated with each of them. A complete knowledge source, using the probabilistic reasoning paradigm, is dedicated to the risk-bearing attitude. After the consultation, selected risk measures are evaluated and posted to the global blackboard (where they are asserted as simple Prolog/Rex facts) to be used by the knowledge source performing MCDM (operational research paradigm). The explicit consideration of multiple, even conflicting, objectives in a decision model has made the area of Multiple Criteria Decision Making (MCDM) very popular among researchers during the last two decades. In our work, we adopted a special outranking method, proposed by Brans [3], based on extensions of the notion of criterion. In INVEX, this method is implemented in a knowledge source using the proceduralprogramming paradigm (C language). Its results are posted to the global blackboard, and further passed to Excel (through the DDE) to be presented graphically to the end user. A multiparadigm method adopted in INVEX is somewhat complex, but so is the problem with which it is designed to deal. The major objective of INVEX is accomplished, by making this complex task easier for the user, while still incorporating all the relevant factors that are critical for the VRANES AND STANOJEVIC: PARADIGMS WITHIN THE BLACKBOARD FRAMEWORK decisionmaker in the real world. INVEX has confirmed the validity of the assumption [20] that knowledge-based computing cannot stand apart from conventional data-processing techniques and concerns, and most future complex applications will be only partially knowledge-based, with the remainder of the system being built out of conventional technological components. VI. CONCLUSION This paper reports one possible way of integrating diverse knowledge representation and reasoning techniques within knowledge-based programming paradigms (internal integration) and the integration with other programming paradigms (external integration), using the global blackboard as an integration platform. The BEST-based multiparadigm system comprises several single-paradigm knowledge sources, each performing the function for which it is best suited, communicating via abstract data types representing the common substrates on the global blackboard. The combination of multiple programming paradigms supplies a firm foundation for addressing complex problems such as process control, command and control, planning and resource allocation. A multiparadigm approach to complex problem solving is .much more adaptable and thus less brittle. Moreover, a multiparadigm approach solves the problem in a more natural way, increases programmer productivity during application development, reduces maintenance costs, and significantly improves the overall efficiency of the system, since every particular subproblem is solved using the most appropriate paradigm. The deep integration of paradigms accomplished through BEST provides all the advantages of all technologies, giving the programmer the right tool at the right time. With our multiparadigm framework we have expanded the range of applications that could exploit intelligent reasoning and knowledge processing, including complex, heterogeneous, evolvable and scalable applications. Moreover, BEST allows the importing of commercial tools (Excel in INVEX for example), letting users develop each subsystem with the best available technique. The system is fully implemented, and has proven to be very suitable for building complex applications, such as those mentioned in Section V. REFERENCES [I] J. F. Allen, “Maintaining knowledge about temporal intervals,” Commun. ACM, vol. 26, no. 11, pp. 832-844, Nov. 1983. [2] R. .I. Brachman, Richard E. Fikes, and Hector J. Levesque, “KRYPTON: A functional approach to knowledge representation,” IEEE Comput., Special Issue on Knowledge Representation, vol. 16, pp. 67-73, Oct. 1983. [3] J. P. Brans and Ph. Vincke, A preference rankina organization method. Munagemenr Sci., vol. 31, no. 6 pp. 647-656, 19851 [4] B. S. Buchanan and E. H. Shortlife, Rule-Based Experr Sysrems. Reading, MA: Addison-Wesley, 1985. [S] A. Cl. Cohn, “A more expressive formulation of many sorted logic,” J. Automated Reasoning, vol. 3, no. 2, pp. 113-200, 1987. [6] J. de Kleer, “An assumption-based truth maintenance system,” Artificial Intell., vol. 28, pp. 127-162, 1986. [7] -, “Extending the ATMS,” Arfijicial Well., vol. 28, 1986, pp. 163-196. “Problem Solving with the ATM&” Art$cial Intell., vol. 28, F31-3 pp. 197-224, 1986. 261 [91 P. Devanbu, M. Freeland, and S. Naavi, “A urocedural aouroach to search control in Prolog,” in Proc. In?. ‘Conf: EkAI’86, 1988.’ [lOI M. Dincbas and J-P. La Paoe. “Metacontrol of logic oronrams in METALOG,” in Proc. In?. C& Sth Generation Coiput: Syst. 1984, 1984. [Ill J. Doyle, “A truth maintenance system,” Artificial Intell., vol. 12, pp. 231-272, 1979. r121R. Enaelmore and T. Morgan, Blackboard Systems. Reading, MA: Addison-Wesley, 1988. [I31 D. W. Etherinaton, A. Boraida, R. J. Brachman, and Henry Kautz, “Vivid knowledge and tractable reasoning: Preliminary repo&,” Proc IIth Inf. Conj Artificial Intell., Aug. 1989, pp. 1146-l 152. u41 A. A. Faustini and E. B. Lewis, “Toward a real-time dataflow language,” IEEE Sojiware, pp. 29-35, Jan. 1986. fast algorithm for the many pattern/many object 1151 C. L. Forgy, “RETE-A pattern match problem,” Art&al Intell., vol. 12, pp. 17-37, 1980.. 1161A. M. Frisch, “A general framework for sorted deduction: Fundamental results on hybrid reasoning,” Ronald J. Brachman, Hector J. Levesque, and Raymond Reiter, Eds., in Proc. Ist Inc. Con& Principles Knowledge Representation Reasoning, Toronto, May 1989. (171 A. Frisch and A. Cohn, “1988 Workshop on principles of hybrid reasoning,” AI Mug., vol. 11, no. 5, pp. 77-84, Jan. 1991. [lsl S. L. Fulton and C. 0. Pepe, “An introduction to model-based reasoning,” AI Expert, pp. 48-55, Jan. 1990. [I91 H. Gallaire and C. Lasserre, “Metalevel control for logic programs” in Logic Programming, K. Clark and Tamlund, Eds. New York: Academic, 1982. F. [201 Hayes-Roth, J. E. Davidson, L. D. Erman, and J. S. Lark, “Frameworks for developing intelligent systems: The ABE systems engineering environment,” IiEI?Experty vol. 6, pp. 3&41, June- 1991. 1211ART-Automated Reasoning Tool User’s Manual, version 3.0, Inference Corp., Los Angeles, CA, 1987. PI KEE Software Development System User’s Manual, IntelliCorp Inc., Mountain View, CA, 1986. r231 M. A. Jenkins, J. I. Glasgow, and C. D. McCrosky, “Programming Styles in Nial,” IEEE Software, pp. 4655, Jan. 1986. New York: Wilev. ~241 S. Khoshatian and R. Abnous. Object Orientation. 1990. ]251 G. J. Klir and T. A. Folger, Fuzzy sets, Uncertainty, and Information. Englewood Cliffs, NJ: Prentice-Hall, 1988. 1261B. Kowalski and L. Stipp, “Object processing for knowledge based systems,” AI Expert, Oct. 1990. ~271 P. B. Ladkin, “The completeness of a natural system for reasoning with time interval,” in Proc. I&h In?. J. Cont. Milan, Italy, 1987, pp. 462-467. @I J. Liebowitz, “Integrating intelligent systems.” Intell. Svst. _ Reu.. _ vol. 10, no. 1, pp. 7-8,-1993. . ~291 G. Lindstrom and P. Pananeaden. “Stream-based execution of logic programs,” in Proc. I984 in?. Simp. Logic Programm., 1984, pp. 168-176. [301 G. Lindstrom, “Functional programming and the logical variable,” in Proc. Symp. Principles Progr-aim. Languages, Jan. 1985, pp. 266280. [311 R. MacGregor and M. H. Burstein. “Using a description classifier to enhance knowledge representation,” IEEE Expert, vol. 6, pp. 4146, 1991. L. J. Mazlack, “Satisfying in knowledge-based systems,” Data & [321 Knowledge Eng., vol. 51 pp. 139-166, 1990. 1331 P. Patel-Schneider. “Decidable first-order logic for knowledge representation,” Ph.D. dissertation, Univ. Toronto, May 1987. Reading, MA: Addison Weslev, 1984. 1341 J. Pearl, Heuristics. 1351 C. Rich and Y. A. Feldman:“Seven layers of knowledge representation and reasoning in support of software development,” IEEE Trans. Sofhvare Eng., vol. 18, pp. 451-469, June 1992. ]361 N. C. Rowe, Artificial Intelligence Through Prolog. Englewood Cliffs, NJ: Prentice-Hall, 1988. [371 M. J. Stefik, D. G. Bobrow, and K. M. Kahn, “Integrating accessoriented programming into a multiparadigm _ - environment,” IEEE Sofiware, voi. 3: Jan.1986. [381 M. E. Stickel. “Automated deducton by theory resolution,” in Proc. 9th In?. Joint Conj Artificial Intell., Los- Angeles, CA, Aug. 1985, pp. 455458. [391 P. Subai, M. StanojeviC, and S. VraneS, “Adaptive control in rule based systems,” IMACS Symp. Modeling and Control of Techn. Syst., Lille, France, May 1991. ]401 M. Tokoro and Y. Ishikawa, “Orient/84: A language with multiple paradigms in the object framework, ” in Proc. Hawaii In?. Conf: Syst. Sci., Honolulu, Hawaii, Jan. 1986. 1411 M. Vilain, ‘The restricted language architecture of a hybrid representation system,” in Proc. In?. Joint Co@ Arti$cial Intell., Morgan Kautman, 262 IEEE TRANSACTIONS Los Altos, CA, 1985, pp. 547-551. ~421 S. VraneS, M. Stanojevic, M. Lu-in and V. Stevanovic, “Blackboard metaphor in tactical decisionmaking,” Europ. J. Operational Rex, vol. 61, no. 1-2, Aug. 25, 1992. [431 S. VraneS, M. Stanoievic, M. Lu-in, V. Stevanovic, and P. Subai, “A blackboard framework on top of prolog,” ht. J. Expert Syst. Applications, vol. 7, no. 1. Dec. ‘93/Jan. ‘94. way to extend prolog [441 S. VraneS and M. Stanojevic, “PROLOG/REX-A for better knowledge representation,” IEEE Trans. Knowledge Data Eng., vol. 6, pp. 22-38. “IN[451 S. VraneS, M. Stanojevic, V. Stevanovic, and M. Lu-in, VEX-Investment advisory expert system,” submitted for Intelligent Systems in Accounting, Finan& and kanagement. [461 C. Walther, A Many-Sorted Calculus Based on Resolution and Paramodulation. Los Altos, CA: Morgan Kaufman, 1987. 1471 P. Zave, “A compositional approach to multiparadigm programming,” IEEE Software, vol. 6, pp. 15-25. Sanja VraneS (M’90) received the Dipl. Ing., MSc., and Ph.D. degrees in electrical engineering from the University of Belgrade, Yugoslavia. She has worked as a Research Scientist at the Mihajlo Pupin Institute for Automation and Telecommunications, Belgrade, since 1980. She took a oneyear sabbatical leave from 1993 to 1994 working at the Advanced Manufacturing and Automation Research Centre (AMARC) at the University of Bristol, England. Her primary research interests are in the fields of multiparadigm programming, knowledge-based systems design, blackboard systems, knowledge representation, truth maintenance, and machine learning, in which she has published more than 40 papers. Dr. VraneS is a member of the IEEE Computer Society and the AAAI. View publication stats ON SOFIWARE ENGINEERING, VOL. 21, NO. 3, MARCH 1995 Mladen StanojeviC received the Dipl. Ing. degree in computer science (cum laude) from the University of Belgrade, Yugoslavia, in 1989. He is working towards the MSc. degree at the same university. He is currently a Graduate Research Assistant in the Computer Systems Department of the Mihajlo Pupin Institute, Belgrade. His research interests include knowledge representation, knowledge-based systems, blackboard systems, truth maintenance systems, object-oriented and logic programming.