A blackboard architecture for plastics design V. Venkatasubramanian and Chun-Fu Chen
Intelligent Process Engineering Laboratory, Department of Chemical Engineering, Columbia Unirersity, New York, N Y 10027, USA Engineering design is a complex, ill-structured problem involving vast amounts of knowledge that often deal with incomplete, and uncertain, information. The design process typically involves a number of sub-tasks, some of which are amenable to numeric or algorithmic processing (e.g., analysis and optimization); most other sub-tasks require symbolic processing, and are typically solved by expert designers who rely on experiential knowledge used in the form of heuristics. The latter set of sub-tasks were chiefly responsible for creating the symbolic bottleneck that has set back the efforts to automate the design process. However, with the advent of recent AI methodologies we now have techniques for tackling the symbolic bottleneck. Even though some progress has been made in this area with the aid of knowledge-based expert systems, many important questions still remain to be researched. In the DESIGNER project we are performing empirical studies that focus on these questions using the plastics design problem as our test bed. We have focused on the preliminary or the base-case design in this domain and have implemented a working prototype called DESIGNER, based on the blackboard architecture. This paper discusses the various aspects of DESIGNER, which is a hybrid, integrated, conferring expert system (HICEX). We are currently extending DESIGNER's capabilities to address the other aspects of the plastics design problem such as part design, plastics processing selection, etc. We are also investigating decentralized communication architectures for integrating multiple experts o n the design problem. Key Words: blackboards, engineering design, plastics, conferring expert systems.
1. NATURE OF E N G I N E E R I N G DESIGN Engineering design is an important, but poorly understood discipline. This is so because it is a complex, inexact decision-making process, largely solved by experimental knowledge, and thus is not amenable to an algorithmic approach. In a chemical process industry, the amount of time, money and effort spent on the design of process equipment, process synthesis, flowsheet preparation, etc. routinely takes up 50 to 60 percent of the resources available 1. With the advent of the AI technology, however, we now have some hope of automating this difficult process which would result in substantial savings in resources. In the DESIGNER project we attempt to understand, from an AI perspective, the generic tasks that may be involved in the design process and thus explore the possibility of developing a domain independent methodology.
1.1. Characteristics of the design process A typical process design problem in the chemical industry (and also in other domains) involves a number of subtasks: •
selection processes such as selections of models, materials, equations and mathematical formalizations, parameters, etc.;
Accepted August 1986. Discussioncloses December 1986.
0267-9264/86/020117-0652.00 © 1986Computational Mechanics Publications
• • • •
synthesis processes such as putting various components, models, or processes together; analyses such as simulation, prediction, optimization, etc.; evaluation of the overall system regarding its performance, in light of problem constraints such as efficiency, cost, safety, etc.; modifications of the sub-components to meet the necessary constraints.
In striving towards the desired solution to a design problem, one often resorts to a hierarchical decision procedure resulting in various levels of abstraction to manage the computational complexity of the search space. One starts with the rough features of the desired solution and by traversing various levels of abstraction one gets more and more specific and detailed about the solution. The subtasks mentioned above are often carried out many times over in reaching the ultimate goal. Of these subtasks, only the analysis part is amenable to algorithmic, numeric processing. This part has been largely automated by the use of computers. The rest of the design process is substantially symbolic processing, involving inexact or probabilistic reasoning, which is highly dependent upon the use of heuristic knowledge 2. This larger part has not been automated, and is usually performed by people, and is therefore expensive. Recent progress in AI indicates that the knowledge intensive operations can be automated by the use of AI techniques 3.
Artificial Intelligence, 1986, Vol. 1, No. 2
117
A blackboard architecture: V. Venkatasubramanian and Chun-Fu Chen 1.2. Towards a computational model of the design process It must be quite clear from section 1.1. that engineering design is a complex process involving many interacting subtasks that are repeated many times over until the final goal is achieved. Design problem-solving demands flexibility in knowledge representation, reasoning strategies and control. The information processing architecture must be able to support a wide variety of knowledge representation schemes such as frames, rules and procedures that are appropriate for different types of knowledge required in design. The architecture must also facilitate a combination of reasoning strategies such as data-driven, goal-driven, opportunistic reasoning, etc., as such a variety of techniques are used by expert designers. One also requires the ability to support multiple levels of abstraction, and efficiently integrate multiple agents or experts. We propose that the complex task of process design cannot be adequately performed by a single expert system, but calls for a group of cooperating expert systems, each an expert in some specific task domain. We refer to this paradigm as the conferring expert systems methodology. Recent experiments with design expert systems at Carnegie-Mellon University also point towards such a formalism4. Another important requirement for design expert systems in process engineering is that they should integrate both the symbolic and the numeric information processing aspects of a design problem. The design system should ultimately have numeric procedures to perform computations such as simulation and optimization, which should be fully integrated with the decision-making aspects of design. One is thus naturally led to Hybrid Integrated Conferring Expert Systems (HICEX). The architectural principles of these systems are not fully understood and several important questions remain to be answered. The purpose of our research is to investigate these questions by experimenting with a number of different process engineering design problems. Some of the important questions that arise in the design of HICEX are: Knowledge representation issues Since we have a group of conferring expert systems, they might have different knowledge representation schemes appropriate to their domains. For example there could be a production rule representation for encoding heuristic knowledge, a frame representation for encoding concepts or objects, procedural representations for performing numerical computations, etc. We then face the question of how to efficiently integrate all these together. Control and Communication How is such a distributed problem solving system to be controlled? How do the various expert modules communicate with each other? Recent advances have focused on the so-called blackboard architecture 4, a model for mediating the communication between the various modules. This leads to a somewhat centralized control mechanism. There are other alternatives for regulating communication using a more decentralized architecture and loosely coupled expert modules following some of the techniques from the organization of communication networks 5.
118
Artificial Intelligence, 1986, Vol. 1, No. 2
•
•
Hierarchical organization How is such a complex system organized vertically? Clearly there is a need for a hierarchical structure which facilitates effective control of reasoning and overall organization of the system. The issue of proper organization knowledge at various levels of abstraction is important for process design. Inference techniques As design demands the use of a variety of inference techniques such as backward and forwards chaining, means-and-end analysis, generate-and-test methods, stochastic relaxation 6, etc., the proper selection and use of these strategies at various times in the design problem-solving is important. One issue that should be looked at is the ability to support and switch among various inference engine modules in a manner similar to engaging various knowledge sources or modules.
In the DESIGNER project we address these and other related issues by working on the problem of plastics design. The selection of an appropriate plastic material for a given use and the specification of its processing to realize the the end product is an important problem in the polymer and plastics industry7. This problem also has the typical features of a design involving the sub-tasks listed in section I.I., thus serving as a nice testbed for performing empirical studies. 2. BLACKBOARD ARCHITECTURE FOR MODELLING ENGINEERING DESIGN The blackboard architecture was originally developed for the Hearsay-II speech understanding system8 as the problem-solving methodology. Considerable work has been donesince then in applying the model to diverse problems such as multiple-task planning 9, protein crystallography ~o, building design ~t, catalyst selection 4, etc. Even though these systems differ from one another substantially in many aspects, deep down they all share the essential features of the blackboard architecture. The diversity of applications evolve from the model's flexibility and generality, as we shall see in subsequent sections (see Ref. 12 for a comprehensive introduction to blackboards).
2.1. Essential features of the blackboard architecture The blackboard architecture has four salient features: (1) entries, which are intermediate results generated in the problem-solving process, (2) knowledge sources, which are independent, data-driven, modules of knowledge that produce entries, (3) the blackboard, a structured global database that maintains the intermediate results of the problem-solving process (entries) and facilitates the indirect interaction among the knowledge sources, and (4) a control mechanism that determines how and when the knowledge sources should be activated and allowed to modify the contents of the blackboard database. These features are discussed in greater detail in the context of our DESIGNER system in the next section. The essentials of a blackboard are so general that many information-processing problems can be easily fitted into this paradigm. This accounts for the flexibility and the diversity of the architecture. Flexibility is also derived from the model's liberal control aspects and knowledge representation requirements. As it must be obvious, the
A blackboard architecture: V. Venkatasubramanian and Chun-Fu Chen narrowing down the alternatives and offering some overall feel for the problem. Knowledge-based expert systems are most appropriate for this preliminary design process and they can serve as valuable assistants to the design engineer. We have focused on this preliminary design process for plastics engineering, and have developed a prototype expert system called DESIGNER that uses the blackboard architecture.
events goals
BtackBoa
td Q
Q
Fig. 1. A schematic of the blackboard architecture, where EKS stands for Expert Knowledge Source, F.A. for Focus of Attention, and U.I. for User Interface blackboard architecture is just a generalized production system model. The blackboard and the entries are the equivalents of the workin9 memory and its contents, namely, the data. The knowledge sources correspond to the production memory except that the production memory has been divided into a number of independent modules. Fig. 1 represents a schematic of the blackboard architecture. 3. D E S I G N E R : A H Y B R I D EXPERT S Y S T E M F O R
PLASTICS DESIGN DESIGNER is a hybrid, integrated, conferring expert system (HICEX) implemented using the essential ideas of the blackboard architecture. It addresses some of the important problems in the plastics processing industry requiring the selection and the design of plastics components given some overall goals. This problem has the characteristic features typical of engineering design. The following sections outline the details.
3.1. Features of the problem The plastic component engineering problem has three primary aspects, namely, selection, design and processing. The selection process typically involves the selection of one or more plastic materials as potential candidates given the overall constraints such as the intented use, cost limitations, etc. The design process involves mechanical design of the component and the processing part is concerned with manufacturing decisions such as what type molding method to use, the important features of the mold, and so on. This complete process (simply called the design process, even though it involves selection as well as processing) is very complex and detailed. However, often one can arrive at rough estimates about the product such as its applicability potential, important design restrictions, processability criteria, and of course, t he cost of manufacture and profitability potential, by performing the so called preliminary design or base-case design. Expert designers often use experience-based heuristics to perform short-cut methods to arrive at some preliminary evaluations about the problem. Even though the conclusions arrived at by this approach are not precise, they are often sufficient to rule out some alternatives v. Thus, the approach saves much time and effort by
3.2. Architecture of DESIGNER Despite the complexity of the design process, the problem-solving can be hierarchically classified, and within the hierarchy one can further divide the task into a number of loosely coupled modules, each of which concentrates on some relatively small subtask. Such a problem-solving process is suitable for the blackboard architecture with the modules serving as expert knowledge sources (EKS). Owing to the hierarchical structure, we are really dealing with a multi-layered blackboard architecture with each layer having its own pool of knowledge sources and a communication medium.
3.2.1. Knowledge representation issues As mentioned in section 1.2., the design process requires a hybrid, integrated, conferring expert system architecture which is facilitated in DESIGNER by using the blackboard model. Knowledge is represented in the form of rules and frames, thereby providing a hybrid fiavour. Frames are generalized property lists and are extremely useful for representing objects with various properties and inter-relationships. In the frames terminology, properties or attributes are represented as slots and the symbolic or numeric values of the attributes are represented by the values of the appropriate slots. Frames also facilitate the transfer of properties and values of one object to another automatically through a process called inheritance where the latter object inherits certain properties (and values) from the former ~3. The plastics are represented as frames with appropriate slots for representing properties. They are organized as a hierarchy of frames, thus facilitating inheritance when needed. Constraints such as the cost of a plastic material or some desirable property such as acid resistance are also represented as frames. The constraints also naturally fit into a constraint hierarchy. Rules are typically used to represent control knowledge. The hybrid composition of rules and frames is facilitated by FranzLisp functions which act an interface. In DESIGNER rules are implemented in OPS514 and frames in Framesmith is. The entire knowledge base is composed of more or less independent, loosely-coupled modules called expert knowledge sources (EKS). Some of these EKS are problem-dependent such as the chemical expert (CHEM_ EKS), and some are problem-independent like the user interface (UI_EKS), focus of attention (FA_EKS), etc. These modules contain collections of rules specific to the use of that EKS, and these rules can communicate with the frame knowledge base through the use of Lisp functions. The CHEM_EKS is an expert knowledge source that 'knows' about the various chemical properties such as acid resistance of all the plastics in the knowledge base and thus will be able to assist in recommending an appropriate plastic given the different chemical constraints that need to be satisfied by the desired plastic.
Artificial Intelligence, 1986, Vol. 1, No. 2
119
A blackboard architecture: V. Venkatasubramanian and Chun-Fu Chen Similar properties and duties exist for the other domainspecific EKS's. The UI_EKS is a problem-independent expert knowledge source that manages the user-interface needs such as help facilities and graphics. The Focus of Attention EKS (FA_EKS) deals with the management of the overall control of the blackboard. Each EKS is typically activated by some 9oal posted on the blackboard that satisfies the pre-conditions of the EKS. When more than one is activated, the focus of attention EKS does the conflict resolution, assigns priorities using the heuristics built into it, and determines which EKS should be asked to go ahead with the problem-solving.
3.2.2. Control aspects Control of the various EKS is mediated by the blackboard with the aid of the user interface EKS, scheduler EKS, and focus of attention EKS. At the start of the session the user communicates his desires to the user interJ'ace EKS which assists him in the initial stages of the problem specification process. The focus of attention EKS then makes the preliminary decomposition of the problem and determines the immediate goals to be pursued. The product specification EKS then determines the relevant constraints and posts them on the blackboard. This will activate one or more of the EKS's, and an appropriate EKS will attend to some subproblem, develop the complete or partial solution, and communicate its results to the blackboard. Thus, control is passed on from one EKS to another through the messages posted on the blackboard and with the aid of fiI_EKS which oversees the control process. Problemsolving by a given EKS is carried out by the execution of various inference rules that manipulate the appropriate contents of the blackboard (entries). For this internal control within an EKS the blackboard is again used as the communication medium. However, this does not affect the other EKS's as the problem-solving is now done within the context of the given EKS, and hence the messages posted do not trigger other EKS's. In this respect, a given EKS uses its own private portion of the blackboard as its workin9 memory, and does not affect the other EKS's. This behaviour justifies the interpretation that the blackboard is a 9eneralized workin9 memory, as mentioned before. 3.2.3. DESIGNER user-interface For any expert system to be useful, it must posses a friendly, transparent user-interface. This need is even more pressing for expert systems in engineering domains, where one deals with various types of objects, interactions, processes, and so on. For such system we advocate the use of visual user-interfaces as people understand visually represented information much better than purely textual information. We have strived towards this goal, and our prototype thus has a visual interface that graphically displays the blackboard and its contents, the different expert knowledge sources in action, etc. as the problem-solving process proceeds. From our demonstrations to a variety of people, mostly without an AI background, it has become apparent that the users' tinderstanding and appreciation of the consultation session are substantially facilitated by the presence of such a visual interface. The visual-interface has three parts, namely a consulation window, a rule window, and a black board window. The consultation session with the
120
Artificial Intelligence, 1986, Vol. 1, No. 2
expert system is carried on in the consultation window where the user is consulted through a question-answer dialogue process. The various rules that fire during consultation are displayed in the rule window, thus making the reasoning of the expert system more transparent. The blackboard window shows the current status of the problem-solving by displaying the contents of the blackboard and the expert knowledge source that has been activated and engaged in the consumption. An example of a consultation session is shown in Fig. 2. The row of boxes labelled as FA, UI, etc. are the various EKSs interacting with the blackboard. The figure indicates that the Specification (SPEC) EKS is currently involved in the problem-solving by responding to the goal posted on the blackboard. The specification EKS generates the various constraints the desired plastic must satisfy by using its built-in heuristic rules. For example, if the usage is as laminates in a chemical environment, the SPEC EKS will automatically generate acid resistance, alkali resistance, and organic solvents resistance as the necessary constraints among other more generic ones like minimum cost. The user may also interactively post some constraints or override some of the suggested ones as this flexibility is often important in design. The system is also capable of explaining, if desired, why some constraints were generated and why some others were not. After the preliminary selection is made, decisions about mechanical design and processing are made. As mentioned earlier, we have concentrated on developing a system that will assist the user in preliminary or base-case design by relying on experiential knowledge in the form of heuristic rules. Hence, the system's recommendations are only approximately correct with the benefit of arriving at conclusions quickly. We are currently working on enhancing the performance of the system. 4. CONCLUSIONS AND FUTURE WORK Engineering design is a complex, ill-structured problem involving vast amounts of knowledge that often deal with incomplete, and uncertain, information. The design process typically involves a number of sub-tasks, some of which are amenable to numeric or algorithmic processing; most other sub-tasks require symbolic processing, and are typically solved by expert designers who rely on experiential knowledge used in the form of heuristics. The la.tter set of sub-tasks were chiefly responsible for creating the symbolic bottleneck that has thwarted the efforts to automate the design process. However, with the advent of recent AI methodologies we now have techniques for tackling the symbolic bottleneck. Even though some progress has been made in this area with the aid of knowledge-based expert systems, many important questions still remain to be researched. These questions need to be resolved before substantial progress is seen towards the automation of engineering design. In the DESIGNER project we are performing empirical studies that focus on these questions using the plastics design problem as our test bed. The problem is a n important one in the plastics engineering industry and any progress towards the design of a computer-based consultation system would be extremely valuable. We have focused on the preliminary or the base-case design in this domain and have implemented a working prototype called DESIGNER based on the blackboard
,4 blackboard urchitecu~re: V. Venkatasubramanian and Chun-Fu Chen 'I:I"~'I~'III'I' L -
1. 2 3, 4.
-----
m i
i
i n
construction packaging transpor tit 1on
elsctronlc
"_-ntor a number:
I
[p sp-flams_product_constructton [messog ^type enterlng_use_tree ^status in process) (product ^category <( t construction ) ) ) --) ( w r i t s ( c r l f ) What is the s p e c i f i c use of your product? ( c r l f ) (to ~to 5) 1. f l o o r i n g ( c r l f ) (tabto 5) 2. Insulation ( c r l f ) (tabto 5) 3.
-~RESS ANY KEY lO CDNTINUEI ~hat is the specific use of your product? I. flooring 2. Insulation
pips ( ¢ r l f ) (tobto 5) 4. siding ( c r l f ) (tabto ~) 5. wall covering (or1 f) Enter • number: ) (bind i x ) (accept)) (remOve 1) (r~nove 2)
3. pipe 4. siding 5. wall covering inter a number: 3
(make product ^c|tmgory construction *use tops~r~ams_product_constr uction ( x ) ) ) )
~RESS ANY KEY TO CONITNUE I
DESIGNER
DEMO
PROTOTYPE
II <= 14 h e s s e 9 t s g a g u s i n _ p r o c e s s t"gTpe e u g e r i n g _ u s e _ g r e e <= 12 p r o d u c t t c a l ; e g o r y 1 => 11 produce fcal;egor7 c o n s t r n c g l o n +use pipe • . 10 9oat 'l'a~z~ns I n _ p r o c e s s ' t a a t ~ s p e c i f y C a s a l | n e d _ g o a p e c l f T _ e k s • " 4 plastics
,,'
tzvallabte
UI
ABS ACETkL PlTL01~ PE POLTCkRBOIIkT[ ?OLTSULFOIE PP PFE PS P¥C
CHEH
[HECB
ELEC
D[Cl
EXPL
Fig. 2. D E S I G N E R ' s user-inter~we showing the consultation window, the rule window, and the blackboard. In the blackboard, = > indicates the most recent entry, < = shows the entries that hat'e been removed in the current cycle, and ** represents past entries
architecture. D E S I G N E R is a hybrid, integrated, conferring expert system (HICEX). The plastics design process requires substantial flexibility in knowledge representation and reasoning as well as the facility to perform hierarchical problem-solving. We found that the blackboard approach to design problem-solving supported these needs effectively. We are currently extending D E S I G N E R ' s capabilties to address the other aspects of the plastics design problem such as part design, plastucs processing selection, etc. We are also investigating decentralized communication architectures for integrating multiple experts on the design problem. 5. A C K N O W L E D G E M E N T S The authors with to thank Dr Rene Banares-Alcantara for many fruitful discussions. They also would like to thank Dr M. L. Bushnell for providing them with the Framesmith language. One of the authors iV. V.) would like to convey his appreciation to Professor Art Westerberg and Dr M. D. Rychener for valuable discussions on engineering design. It is a pleasure to thank Professor D. Sriram for many interesting
discussions on artificial intelligence and the design process.
REFERENCES 1 2
3 4 5 6 7 8
Ludwig. L. L. Applied Process Design jbr Chemical and Petrochemical Plams. Gulf Publishing Co, 1977 Banares-Alcantara. R.,-Sriram, D., Venkatasubramanian. V.. Westerberg, A. and Rychener. M. Knowledge-Based Expert Systems for CAD. Chemical Engineering Progress, September 1985.81(9L 25-30 Waterman, D. ,4 Guide to Expert Systems, Addison-Wesley Publishing Company, Reading. Massuchusetts, USA, 1985 Banares-AIcantara.Rene DECADE: Design Expert Jbr CAtalyst DErelopment. PhD dissertation, Chemical Engineering Department. Carnegie-Mellon University. February 1986 Tanenbaum. A. S. Comp,aer Networks. Prentice-Hall, Englewood Cliffs, New Jersey. USA, 1981 Venkatasubramanian, V. An expert systembased on a stochastic parallel network. Proceedings of the Expert Systems in Gorernment Colffbrence, IEEE, McLean, Virginia. October 1985 Schwartz.S. S. and Goodman. S. H. Plastics Materials and Processes. Van Nostrand Publishing Co.. New York, 1982 Erman.L. D.. London. P. E. and Fickas. S. F. The design and an example use of HEARSAY-Ill. Proceedings of the Serenth International Joint Coq]&ence oil Artificial Intelligence, 1981. 409-415
Artificial Intelligence, 1986, Vol. 1, No. 2
121
A blackboard architecture: V. Venkatasubramanian and C h u n - F u Chen 9
10 11 12
122
Hayes-Roth. B.. Hayes-Roth. F., Rosenschein. S. and Cammarata, S. Modelling planning as an incremental. opportunistic process. Proceedinys of the Internatiomtl Joint Cop!t~rence opl ,4rtitk'ial Intelligence. 1979. 375-383 Terry, A. Hierarchical Control of Prod,wtion Systems. PhD dissertation. Department of Computer Science, UC Irvine, 1983 Sriram, D. DESTINK A Knowledge-based approach to integrated structural design. PhD dissertation, Department of Civil Engineering, Carnegie-Mellon University. March 1986 Hayes-Roth, Barbara, The Blackboard Architecture: A General Framework for Problem Solving?, Heuristic Programming
Artificial Intelligence, 1986, Vol. 1, No. 2
13 14
15
Project Report No. H P P-83-30, Computer Science Department. Stanford University. May 1983 Winston. P. H. and Horn, B. K. P. LISP. Addison-Wesley Publishing Co., Reading, Mass.. 1984 Brownston, L., Farrel, R., Kant. E. and Martin, N. Programmin9 Expert Systems in OPS5. An Introduction to Rule-Based Programming, Addison-Wesley Publishing Company, Inc.. Reading. Mass., 1985 Bushnell. Michael L. Department of Electrical Engineering. Ttre Framesmith Manual, Carnegie-Mellon University. PA, USA, 1985, Carnegie-Mellon limited distribution version