16
Pergamon PII:S0952-1976(96)00056-5 EngngApplic. Artif. Intell. Vol. 9, No. 6, pp. 611-626, 1996 Copyright © 1996 IJCAI Inc. Published by Elsevier Science Ltd Printed in GreatBritain.All rightsreserved 0952-1976/96$15.00+0.00 Contributed Paper The Shared Expertise Model for Teaching Interactive Design Assistants TOMASZ DYBALA George Mason University, U.S.A. GHEORGHE TECUCI George Mason University, U.S.A. and Romanian Academy, Romania HADI REZAZAD George Mason University, U.S.A. and ORCHID Technologies and Management, U.S.A. (Received April 1996) The growing complexity of contemporary engineering designs requires the use of sophisticated computer-based design tools. Such tools increase productivity in drafting, configuration and calculations. However, the current generation of design tools plays a rather passive role in the entire design process. Recent progress in knowledge-based engineering design, machine learning and knowledge acquisition allows the development of knowledge-based design assistants which could behave as active partners to human designers, rather than as passive graphical or computational tools. This paper presents the shared expertise model (SEM) of interaction between a human designer and a knowledge-based design assistant, in which the design assistant behaves as an apprentice and a collaborator in the design process. The human designer and his computer-based assistant create designs together, with the assistant proposing routine or even innovative designs, and the human designer correcting and finalizing these designs, as well as specifying creative designs. In this process, the assistant also learns from the human designer, constantly extending and improving its knowledge base, and becoming a better design assistant. This is achieved by employing apprenticeship multistrategy learning based on a plausible version space representation. Within the SEM framework, a human designer with limited programming capabilities can directly develop and maintain a personalized knowledge-based design assistant. The use of SEM leads to increased capabilities of the expert-assistant design team. In this paper, some of the main features of the SEM model, as well as its implementation with the Disciple toolkit, are illustrated within the domain of computer workstation configuration. Copyright © 1996 IJCAI Inc. Published by Elsevier Science Ltd Keywords: Computer engineering, engineering design, configuration problems, intelligent design assistants, apprenticeship learning, multistrategy learning, plausible version space. I. INTRODUCTION The main goal of this research is the development and application of a theory, methodology and toolkit for Correspondence should be sent to: Dr Tomasz Dybala, Learning Agents Laboratory, Department of Computer Science, George Mason Uni- versity. 4400 University Drive, Fairfax, VA 22030, U.S.A. E-mail: [email protected]. This paper has previously been published in Proceedings of the International Joint Conference on Artificial Intelligence. Copyright International Joint Conferences on Artificial Intelligences, Inc. Copies of this and other IJCAI Proceedings are available from Morgan Kaufmann Publishers, Inc., 340 Pine Street, 6th Floor, San Francisco, CA 94104, U.S.A. [http'//www.mkp.com]. building interactive learning agents for complex real-world domains. By a learning agent is meant a specialized interactive knowledge-based system that can be taught by a user to assist him or other users in various ways, for instance, by helping the user to perform his tasks, by performing tasks on the user's behalf, by monitoring events or procedures for the user, by advising other users on how to perform a task, by training or teaching other users, or by helping different users to collaborate. The agent building approach described here, called Disciple, is based on three significant developments in artificial intelligence: apprenticeship learning, ~-3 multi- strategy learning, 4 and programming by demonstration. 5 It 611

The shared expertise model for teaching interactive design assistants

Embed Size (px)

Citation preview

Page 1: The shared expertise model for teaching interactive design assistants

P e r g a m o n

PII:S0952-1976(96)00056-5

EngngApplic. Artif. Intell. Vol. 9, No. 6, pp. 611-626, 1996 Copyright © 1996 IJCAI Inc. Published by Elsevier Science Ltd

Printed in Great Britain. All rights reserved 0952-1976/96 $15.00+0.00

Contributed Paper

The Shared Expertise Model for Teaching Interactive Design Assistants

TOMASZ DYBALA George Mason University, U.S.A.

GHEORGHE TECUCI George Mason University, U.S.A. and Romanian Academy, Romania

HADI REZAZAD George Mason University, U.S.A. and ORCHID Technologies and Management, U.S.A.

(Received April 1996)

The growing complexity of contemporary engineering designs requires the use of sophisticated computer-based design tools. Such tools increase productivity in drafting, configuration and calculations. However, the current generation of design tools plays a rather passive role in the entire design process. Recent progress in knowledge-based engineering design, machine learning and knowledge acquisition allows the development of knowledge-based design assistants which could behave as active partners to human designers, rather than as passive graphical or computational tools. This paper presents the shared expertise model (SEM) of interaction between a human designer and a knowledge-based design assistant, in which the design assistant behaves as an apprentice and a collaborator in the design process. The human designer and his computer-based assistant create designs together, with the assistant proposing routine or even innovative designs, and the human designer correcting and finalizing these designs, as well as specifying creative designs. In this process, the assistant also learns from the human designer, constantly extending and improving its knowledge base, and becoming a better design assistant. This is achieved by employing apprenticeship multistrategy learning based on a plausible version space representation. Within the SEM framework, a human designer with limited programming capabilities can directly develop and maintain a personalized knowledge-based design assistant. The use of SEM leads to increased capabilities of the expert-assistant design team. In this paper, some of the main features of the SEM model, as well as its implementation with the Disciple toolkit, are illustrated within the domain of computer workstation configuration.

Copyright © 1996 IJCAI Inc. Published by Elsevier Science Ltd

Keywords: Computer engineering, engineering design, configuration problems, intelligent design assistants, apprenticeship learning, multistrategy learning, plausible version space.

I . I N T R O D U C T I O N

The main goal of this research is the development and application of a theory, methodology and toolkit for

Correspondence should be sent to: Dr Tomasz Dybala, Learning Agents Laboratory, Department of Computer Science, George Mason Uni- versity. 4400 University Drive, Fairfax, VA 22030, U.S.A. E-mail: [email protected].

This paper has previously been published in Proceedings of the International Joint Conference on Artificial Intelligence. Copyright International Joint Conferences on Artificial Intelligences, Inc. Copies of this and other IJCAI Proceedings are available from Morgan Kaufmann Publishers, Inc., 340 Pine Street, 6th Floor, San Francisco, CA 94104, U.S.A. [http'//www.mkp.com].

building interactive learning agents for complex real-world domains. By a learning agent is meant a specialized interactive knowledge-based system that can be taught by a user to assist him or other users in various ways, for instance, by helping the user to perform his tasks, by performing tasks on the user's behalf, by monitoring events or procedures for the user, by advising other users on how to perform a task, by training or teaching other users, or by helping different users to collaborate.

The agent building approach described here, called Disc ip le , is based on three significant developments in artificial intelligence: apprenticeship learning, ~-3 multi- strategy learning, 4 and programming by demonstration. 5 It

611

Page 2: The shared expertise model for teaching interactive design assistants

612 TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

also emerges from earlier work on the integration of machine learning and knowledge acquisition. 2'6-9 This paper presents the application of the Disciple approach to the development of an engineering design assistant.

The growing complexity of contemporary engineering designs requires the use of sophisticated computer-based design tools. Such tools increase productivity in drafting, configuration and calculations. However, the current gen- eration of design tools plays a rather passive role in the entire design process. In contrast, a Disciple-based design assistant behaves as a collaborator to the human designer. This is very useful because in many real-world domains, and engineering design domains in particular, the space of potential problems to be solved is almost infinite, and the number of parameters to be considered is very large. The designer and his computer-based assistant could interact to create designs. Routine designs would be mostly created by the design assistant under the supervision of the human designer. On the other hand, non-routine designs would be mostly created by the human designer.

In this process, the design assistant also learns from the human designer, constantly extending and improving its knowledge base, and becoming a more useful collaborator. The level of interaction between the designer and the design assistant depends on the quantity and quality of the assistant's knowledge base (KB). Initially, the assistant will behave as a novice that is unable to compose most of the designs. In this case the designer takes the initiative and the assistant learns from the designer. Gradually, the assistant learns to create most of the routine designs within an application domain. Early examples of such design assis- tants are, LEAP-VEXED, ~ Disciple and NeoDisciple. 2'7J°

The engineering design assistant presented in this paper is also an example of an interactive learning agent. This paper will emphasize the shared expert ise model ~'j2 of interaction between the designer and the assistant which allows both:

• the creation of designs through cooperation between the human designer and his knowledge-based design assis- tant, and

• the improvement of the knowledge-based assistant

through apprenticeship multistrategy learning from the human designer.

The rest of this paper is organized as follows. Section 2 contains a general presentation of SEM. Section 3 describes the configuration problem, the design process, and the knowledge representation used. Section 4 contains a more detailed presentation of SEM, and Section 5 gives examples of the assistant's learning capabilities. The last section discusses some of the strengths and weaknesses of the proposed model and presents future research directions.

2. GENERAL PRESENTATION

The methodology for building and using a Disciple-based assistant is illustrated in Fig. 1. First, the agent developer customizes the Disciple Toolkit and develops domain- dependent software modules. For instance, in the case of the Design Assistant presented in this paper, one has to develop the Propose and Revise Design Engine, as well as a specialized interface between the assistant and the designer. The result of this phase is an agent shell that can interact with the expert during teaching and problem solving.

During agent teaching the designer first develops the agent's initial KB, and then teaches the agent how to perform typical tasks. An illustration of this process is given in Section 5. As a result, a minimally knowledgeable agent is developed, that can assist the user in task performance (such as configuring a computer workstation). During agent teaching or problem solving, a need for a more substantial change of the agent may arise, in which case the development phase is resumed.

The Disciple Toolkit is being developed to be used with this agent-building methodology. The toolkit has a modular architecture (see Fig. 2) that facilitates the development of customized agents for various domains. The Disciple Toolkit consists of tools for:

• defining and modifying agent's knowledge, such as the Concept Editor, the Rule Editor and the Dictionary Editor;

Mutual Support

Teaching and Devel~ment ~ " - ~ b l e n ~ Solving

~DISCIPLETOOLKIT / I LEARNING~

I DOMAIN DEPENDENT [ . . . . . . [ A S S IS TANTI WFTWARE MODULES~k~:~edvteO(opmen t y .~/

Fig. 1. The methodology for building interactive learning assistants.

Page 3: The shared expertise model for teaching interactive design assistants

TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL 613

Fig. 2. The architecture of the Disciple toolkit.

• examining the agent's knowledge, such as the Dictionary Browser, the Concept Browser, the Rule Browser, and the Association Browser;

• interactively learning rules, such as the Rule Formulation Tool, the Rule Refinement Tool, the Example Editor, the Example Generator, the Explanation Generator, the Explanation Editor, and the Exception Handler; and

• solving domain problems or performing domain-specific tasks, such as the performance tools.

All the tools rely on the services provided by the Knowledge Query Language. The Knowledge Query Lan- guage provides access to the KB of the agent in a format suitable for learning, design problem solving, and commu- nication with the human designer. This is achieved by employing a heterogeneous knowledge representation which integrates a semantic network representation with a rule-based representation.

A customized agent that behaves as a personal design assistant is composed of the following modules, indicated in the right-hand side of Fig. 3: design knowledge base, propose and revise design engine, learning engine, and GUI-based interface.

The performance modules of the design assistant are based on an interactive propose and revise design method that also facilitates knowledge acquisition and learning. This method is briefly described in Section 3.2.

The main feature of the design assistant is that it is under the constant supervision of a user who is both the developer of the agent as well as the beneficiary of the agent's services. The agent is continuously supervised and taught by the user according to the changing practices in the user's domain, as well as the requirements and the preferences of the user.

The shared expertise model has been proposed to ease and speed up the communication between a human designer and his knowledge-based assistant. An underlying benefit of this cooperative design process is that the assistant acquires new knowledge from the human designer, and by doing so, it improves its design capabilities. An opportunity for the assistant to acquire new knowledge occurs when it encoun- ters a new design which cannot be composed by the assistant because of lack of appropriate knowledge.

An important aspect of the shared expertise model is that the human designer is an integral part of it. The designer's role is two-fold. As a beneficiary of the assistant's services, the designer provides design specifications to be elaborated

Fig. 3. Designer--assistant interactions.

Page 4: The shared expertise model for teaching interactive design assistants

614 TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

by the agent. As a tutor of the assistant, the designer may need to both help and teach the assistant how to compose non-routine designs for which the assistant does not have enough knowledge. The human designer may also need to introduce new knowledge into the assistant's knowledge base, answer queries generated by the assistant, and accept or reject the actions proposed by the assistant. Thus, the cooperation between the human expert and the assistant is carried out, both with respect to the design process and with respect to the learning and knowledge-acquisition process.

The next section focuses on the description of the configuration problem, and on the knowledge representation which facilitates cooperation, based on the design's diffi- culty. Section 4 discusses the algorithm which controls this type of cooperation between the expert and the assistant.

3. INTERACTIVE ENGINEERING DESIGN

3.1. The configuration problem

Engineering design is a process by which products or systems are created to perform desired functions. 13 A configuration problem, so often encountered in contempo- rary engineering design, is defined as designing a product or a system from ready-to-use components. The result of the design activity is a design description. The purpose of such a description is to represent sufficient information about the artifact so that it can be constructed. There seems to be a general agreement on the classification of designs into routine, innovative, and creative categories. TM

The process of designing configurations of computer systems ~5 is an example of a configuration problem. Contemporary computing devices are highly customized with respect to the needed memory, computing power, disk space, input/output devices, etc. The space of all possible configurations is very large. The knowledge needed to make such configurations is mainly heuristic, but it also includes procedures to calculate quantitative design parameters. Most of this knowledge is embedded only in the minds of the human designers. The role of the design assistant built with the Disciple toolkit is to support a human designer in the configuration process, and to build a model of his design knowledge.

An input to the configuration process is a functional specification of an artifact to be designed. This specification characterizes the distinguished functional units of the

product and their features. For instance, Table 1 shows the functional specification of a computer workstation which needs to be configured for a moderate power user who will primarily perform text processing and moderate graphics processing, and will need network communication capabil- ities.

The main task of the design process is to convert the input functional specification into an operational description of the designed product. This describes the configuration of the final product in terms of partial configurations, connections among them, and lists of components used to create these configurations. For instance, the functional requirements of the product described in Table 1 are satisfied by a Macintosh Quadra 840av workstation in the configuration described in Table 2. The workstation is composed of sub-configura- tions, which are described in terms of the components used and their quantities.

3.2. The configuration proeess

There is an inherent order in which designers approach a configuration problem. Typically, they start with an analysis of the functional specifications of a product, acquire new knowledge if necessary, synthesize one or more potential designs, choose one of them, and then evaluate the design constraints based on the original problem specification. If the configuration is not satisfactory, the entire process is repeated.

The cooperation between the designer and the assistant is carried out during these phases of the design process. The analysis is made by the assistant, but the designer serves as a source of the needed information. The synthesis is made by the assistant or by the designer, depending on the problem's complexity. The evaluation is made by the designer.

The learning capabilities of the Disciple-based design assistant will be illustrated by using an example of memory configuration which is a part of the computer-workstation configuration process. Table 3 illustrates the heuristic knowledge used by a system integrator to configure the memory of a computer workstation. Section 5 shows how the design assistant is able to learn such heuristic knowledge.

The configuration of a workstation memory is an illustrative example of the propose and revise design methodJ 6 According to this method the design synthesis is

Table 1. Functional specification of a workstation

?c01

?c02

?m01

is fPorOduct

r

is

is activity

configure ?c02 ?m01

; the goal is to configure a computer workstation ; for a moderate user who will perform ; text processing, moderate graphics ; and network communication

computer-workstation

moderate-user text-processing moderate-graphics communication

Page 5: The shared expertise model for teaching interactive design assistants

TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

Table 2. Operational description of the workstation specified in Table 1

615

?c02 is cat-number memory hard-disk monitor quantity

?s01 is cat-number capacity quantity

?s02 is cat-number capacity access-time quantity

?a01 is cat-number size quality quantity

quadra-840av ; the configured product is a Quadra 840av M6710LLA , with catalogue number M6710LLA; ?s01 ; it is composed of memory, ?s02 ; hard disk, ?a01 ; and monitor;, 1 ; one item of this type is needed.

SIMM-8MB-60NS-32B-72P 11027 8-MB 4

; the memory is configured ; of 4 SIMM memory modules, ; catalogue number 11027; ; each module has 8 MB capacity.

SCSI-disk intemal-component 525-MB 13ms 1

; the hard disk with an SCSI controller; ; it is an intemal, preinstalled component ; with capacity 525 MB ; and access time 13 ms; ; one item of this type is needed.

apple-17-monitor M2612LLA 17" moderate 1

; the monitor is the Apple 17" display ; with catalogue number M2612LLA;

it has a moderate quality resolution; one item of this type is needed.

carried out through a sequence of propositions and revisions of the values of the design parameters. A value for each design parameter is proposed, based on the input functional specification and the values of other design parameters. Once proposed, the value is verified against the design constraints, and possibly revised to maintain design consistency.

The propose and revise design method creates a validated design extension in two steps:

• PROPOSE design extension--a value for a design parameter is either determined by evaluating a procedure

which calculates the value, or a component name selected from a set of available components;

• REVISE design extension additional constraints on the proposed value are checked. If constraints are violated, the proposed design extension is retracted, and revisions to it and/or to other parameters are made. The crucial information to be included in this description is the specification of the parameter to change, how to change it, and some idea of the expert's preference for this revision over others that may be tried.

For example, the memory of a workstation is configured as follows:

Table 3. Heuristic knowledge used to configure the memory of a workstation

• the random access memory (RAM) of contemporary computer workstations is composed of memory banks;

• a memory bank contains a certain number of RAM slots which can be populated with memory modules;

• memory modules can be plugged into memory slots that have the same number of pins;

• the memory module can be either Single Interface Memory Module (SIMM) or Dual In-line Memory Module (DIMM);

• the workstation can accept either SIMMs or DIMMs, but not both;

• the memory modules placed in one memory bank should be of the same type (access time, word length) and capacity;

• memory modules in different banks can be of different types or capacity;

• the total number of the memory modules placed in the RAM slots of all the memory banks give the required memory capacity which should be within a required range;

EM! 9-6-B

Page 6: The shared expertise model for teaching interactive design assistants

616 TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

• propose memory modules which fit into the workstation's memory slots;

• calculate the resulting capacities of the memory banks and the total memory capacity;

• check constraints related to the minimum and maximum memory required, and propose memory modules with different capacities if the constraints are violated.

3.3. The configuration knowledge To transform a functional specification into an opera-

tional description, the assistant needs to have two types of knowledge: descriptions of components and heuristic design rules. The components are described as objects with features, and are hierarchically organized in abstract classes. This hierarchical organization of objects provides the generalization language for learning. For example, the description of MAC Quadra 840av workstation and its relative place in the hierarchy are shown in Fig. 4.

The heuristic rules are if-then rules that describe the conditions under which the steps of the propose and revise design method are performed. The "propose" rules usually indicate the applicability conditions for procedures which calculate the values of some design parameters. A typical procedure retrieves information about relevant decisions made, and initial conditions, and uses mathematical expres- sions to calculate a parameter value. Another type of

"propose" rule is a rule which indicates the applicability conditions for the selection of components that create partial or final configurations. The "revise" rules are fired when values of the design parameters violate some design constraints or the initial specification. They retract values assigned to the design parameters, and replace them with other values that improve design consistency.

A rule may have an exact applicability condition, or two conditions which define a plausible version space (PVS) 6']7 for the exact condition. The two conditions are called the plausible upper bound (PUB) and the plausible lower bound (PLB). The PUB is an expression which is supposed to be more general than the exact condition, and the PLB is an expression which is supposed to be less general than the exact condition. The rule can be fired if one of the two conditions is satisfied. If the rule is fired because the PLB is satisfied, then the consequent of the rule is considered to be true. If the rule is fired because the PUB is satisfied, then the consequent of the rule is considered to be plausible.

An example of a rule with two conditions is informally described in Table 4. The top-right pane of Fig. 5 contains the formal description of the same rule, except that the first line of each bound (?R34) is only partially displayed. This is a "revise design extension" rule. When fired, this rule makes revisions of the design parameters describing the memory module selected for the first memory bank, and the resulting capacity of the first memory bank.

~-7100

;-6100

~3~-CD-RON

0UP, DRA-840AV IS-A O ~ CPU-COMrI6 0N-B0~RD-CPU CPU-PROCESSOR X68040 GPU-CLOC~ 40 PJ~- 0N-B0a.I~ SIIO(-B0~KD- 81~ RMI-X~X-COHFIG 128 R/~-XZH-C01~I6 8 ~ - A C C [ S S - T I ) [ 60-HS R,MI-Bb.~-SI.0TS 1 -SLOe P,~I-SLOTS 4-SLOTS

)UtC-II I~C-LC pO~t'~-BOOK

PS2-?0 Q ~

B U S - ~

IlS -kCCZSS-~'ZXE

l l l a J ~ - OP-P131S

Fig. 4. Concept browser and concept editor.

Page 7: The shared expertise model for teaching interactive design assistants

TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL 617

The lower-left pane o f Fig. 5 shows an instantiation o f the rule that should be interpreted as follows:

The p r o b l e m : revise design extension of a Macintosh model MAC-Ilvx that should have a maximum memory capacity of 32 MB, and the design extension is described by the following

parameters: ---the memory module selected for the first memory bank

is SIMM-16MB-80NS-8B-30P, - - the capacity of the first memory bank is 64 MB, and MAC-Ilvx contains a SIMM-SLOT-30P and

Table 4. Informal description of a PVS rule

Plausible Upper Bound IF:

the problem is to revise a design.extension of a partially configured computer (?M36) that should have a maximum memory capacity (max-mem-l,mit) of the order of MegaBytes (?X43), and the design extension is described by the following parameters:

- the memory module selected for the first memory bank (mm-in-bankl), - the capacity of the first memory bank (bank-l-capacity),

and the capacity of the first memory bank (?X44) is greater than the maximum memory capacity (?X43)

and the computer contains an expansion slot (?$64)

and there is a memory module (?$37) that can be plugged into the expansion slot (?$64), and its capacity is of the order of MegaBytes (?X62)

and there is another memory module (?$41) that can be plugged into the same expansion slot (?$64), and its capacity is of the order of MegaBytes (?X63)

and the capacity of the second memory module (?X63) is less than the capacity of the first memory module (?X62)

Plausible Lower Bound IF:

the problem is to revise a design extension of a partially configured workstation (?M36) that should have a maximum memory capacity (max-mem-limit) of 32 MegaBytes (?X43), and the design extension is described by the following parameters:

- the memory module selected for the first memory bank (mm-in-bankl), - the capacity of the first memory bank (bank-1 -capacity),

and the capacity of the first memory bank (?X44) which is 64 MegaBytes is greater than the maximum memory capacity (?×43) which is 32 MegaBytes

and the workstation contains a RAM memory slot (?$64)

and there is a RAM memory module (?$37) that can be plugged into the RAM memory slot (?$64), and its capacity is of the order of MegaBytes (.9)(62)

and there is another RAM memory module (?$41) that can be plugged into the same RAM memory slot (?$64), and its capacity is of the order of MegaBytes (?X63)

and the capacity of the second memory module (?X63) is less than the capacity of the first memory module (?X62)

THEN

retract the values of the following parameters of the designed artifact (?M36): - the memory module selected for the first memory bank (mm-in-bankl), - the capacity of the first memory bank (bank-1 -capacity),

and extend the designed artifact (?M36) with the following parameter description:

- the parameter =memory module selected for the first memory bank (mm-in-bankl)" has a value described by the variable ?$41.

Page 8: The shared expertise model for teaching interactive design assistants

618 TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

the SIMM-16MB-80NS-8B-30P memory module with capacity 16 MB can be plugged into the SIMM-SLOT- 30P memory slot.

has the solut ion: retract the values of the following parameters of the MAC-Ilvx Macintosh: -- the memory module selected for the first memory bank

(mm-in-bankl), - - the capacity of the first memory bank (bank-

1-capacity),

and extend the designed MAC-Ilvx Macintosh with the following parameter description: -- the "memory module selected for the first memory

bank (mm-in-bankl )" is SIMM-8MB-8ONS-8B-3OP.

Because this solution is generated by using the PLB condition of the rule, it is considered correct.

However, if the problem were to revise the same parameters for a file server which cannot be classified as a

RULE-PI-3-30 RULE-PI-4-2 RULE-P 1-4 - 3 RULE-P1-4-4 RULE-PI0-4-5 RULE-P2-3-30 RULE-P2-4-~ RULE-P2-4-3 RULE-P2-4-4 RULE-P20-3-30 RULE-P22-3-30 RULE-P3-4-3 ?R34 RULE-P3-4-4 .~X43 RULE-P4-4-3 ?X44 RULE -P4 -4 -4 ?M36 R U L E - P S - 4 - 3 ?$37 RULE-R1-4-10 ?$41 RULE-R2-4-10 ?X63 RULE-B3-4-10 ?X62

?S64 RULE-RS-4-10 RULE-R6-4- I0 RETRACT

?P38 EXTEND -BY ?P39

P l a u s i b l e U p p e r B o u n d I F ?R34 REVlSE-DESION-EXTENSION , ARTIFACT-DESIGN ?M36 , MAX-MEM-LIMIT ?X43 , ?X43 MB-CAPACITY ?X44 MB-CAPACITY , OREATER-THAN ?X43 7M36 COMPUTER, CONTAINS ?S64 7537 MEMORY-MODULE, PLUGS-INT0 ?S64, MEMORY-CAPACITY ?X62 ?S41 MEMORY-MODULE , PLUOS-INT0 ?$64 , MEMORY-CAPACITY ?X63 ?X63 MB-CAPACITY , LESS-THAN ?X62 ?X62 MB- CAPACITY ?$64 E)~ANSION-SLOT

Plausible Lower Bound IF REVISE-DESIGN-EXTENSION , ARTIFACT-DESIGN ?M36 , MAX-MEM-LINIT ?X43 , 32 -MB 64-MB , 8BEATER-THAN ?X43 WORKSTATION , CONTAINS ?$64 RAM-MODULE , PLUOS-INT0 ?$64 , MEMORY-CAPACITY ?X62 RAM-MODULE , PLUOS-INT0 ?$64 , MEMORY-CAPACITY ?X63 MB-CAPACITY , LESS-THAN ?X62 MB-CAPACITY RAM-SLOT

PARAMETER ?P38 , ARTIFACT-DESIGN ?M36 PARANETER-DESCRIPTIQN , MM-IN-BANK1 7537 , BANK-I-CAPACITY ?X44 PARAMETER ?P39 , ARTIFACT-DESIGN 7M36 PARAMETER-DESCRIPTION , MM-IN-BANKI ?$41

INPUT Problem / Situation ?R34 REVISE-DESIGN-EXTENSION , ARTIFACT-DE 64-M8 8BEATER-THAN 32-MB ?1436 MAC-II%"A , CONTAINS 7564 ?$37 SIMM-16MB-80NS-SB-30P , PLUOS-INT0 ?S ?$64 $IMN-SLOT-30P

OUTPUT Solution / Action RETRACT PARAMETER ?P38 , ARTIFACT-DESIGN ?1136 ?P38 PARAMETER-DESCRIPTION , MM-IN-BANK1 ? EXTEND-BY PAR~3~TZR ?P39 , ARTIFACT-DESIGN ?M36 ?P39 PARAMETER-DESCRIPTION , ~Qf-IN-BANK1 ? ?$41 SIMM-SMB-80NS-8B-30P , PLUGS-INT0 ?$6 8-MB LESS-THAN 16-MB

INPUT P~oblem / Situation ?R34 REVICE-DESIGN-EXTENSION , ARTIFACT-DE 64-MB 8REATER-THAN 32-MB ?M36 MAC-IIVX , CONTAINS ?$64 ?$37 SIMM-16MB-80NS-8B-30P , PLUOS-INT0 ?S ?$64 SIMM- SLOT-30P

OUTPUT Solution / Action RETRACT P ~ ?P38 , ARTIFACT-DESIGN ?M36 ?P38 P~-DESCRIPTION, MM-IN-BANK1 ? EXTEND-BY PARAMETER ?P39 , ARTIFACT-DESIGN ?M36 ?P39 P~-DESCRIPTZDN , MM-IN-BANKI ? ?$41 SINM-8MB-80NS-SB-30P , PLUSS-INT0 ?$6. 8-MB LESS-THAN 16-MB

Fig. 5. Rule browser.

Page 9: The shared expertise model for teaching interactive design assistants

TOMASZ DYBALA et al.: SHARED 619 EXPERTISE MODEL

ADS Assistant Design Space

TDS Target Design Space IDS Innovative Design Space

RDS Routine Design Space

Fig. 6. Types of design spaces.

workstation, but can be classified as a computer, then the PLB condition would no longer be satisfied. However, the PUB condition would still be satisfied and the solution indicated by the rule would be considered only as plausible.

Such PVS rules are learned by the design assistant from the expert designer by applying the PVS apprenticeship learning algorithm which is discussed in Section 5. During learning, the two bounds progressively converge toward the exact applicability condition of the rule. However, due to the incompleteness and the partial incorrectness of the assistant's knowledge, there is no guarantee that the two bounds will become identical. This is not a weakness of the model. On the contrary, it allows the agent to perform plausible inferences and to continuously improve its knowledge, as will be shown in the remaining sections of the paper. Moreover, this type of rule representation supports a natural cooperation between the expert and the assistant in design and learning. The use of PVS rules for interactive engineering design is controlled by the algorithm discussed in Section 4.

The KB of the assistant is incomplete, and possibly partially incorrect, because the objects may be incompletely defined and the rules (especially those with two conditions) may be only approximately correct. However, during interactive design, the KB is continuously improved to become more complete and more accurate. The plausible version space representation of the rules allows the assistant to perform both deductive and plausible reasoning in a very flexible design space as discussed in the next section.

4. S H A R E D E X P E R T I S E

4.1 . D e s i g n s p a c e s

Figure 6 shows various design spaces of the Disciple- based design assistant.

Consider a design assistant with a complete and correct KB. The set of designs which could be constructed using such a hypothetical KB is called the target design space (TDS). An expert's design space is a good approximation of TDS. The learning goal of the design assistant is to improve its KB so as to approximate, as accurately as possible, a complete and correct KB.

The typical KB of a design assistant is most likely incomplete and partially incorrect. All the design descrip- tions that could be constructed by using the assistant's representation language represent the assistant design space (ADS).

The routine design space (RDS) represents the set of designs that can be deductively constructed using the current KB. Deductive derivations are made when exact rule conditions or PLB rule conditions are satisfied when the rules are applied. As indicated in Fig. 7, some deductively constructed designs are wrong (those within the set RDS- TDS), because the current KB is partially incorrect. Other correct designs are not deductively derivable (those within the set TDS-RDS), because the KB is incomplete. Designs composed within RDS are called routine designs (DR E RDS).

The innovative design space (IDS) represents the set of designs that can be composed from the KB by using either deductive or plausible inferences. Plausible inferences are made by using the rules from the KB when only their PUB conditions are satisfied. Designs composed by making at least one plausible inference are called innovative designs (DI ~ IDS-RDS). The design assistant can create an innova- tive design DI based solely on the existing KB and its plausible reasoning capabilities.

Designs composed outside IDS but within TDS are called creative designs (DC e TDS-IDS). A creative design DC is composed by the expert designer who introduces new object descriptions and design parameters and therefore extends

- ADS, target designs not definable in ADS

n TDS - IDS, target designs not defined in ADS "~ TDS, correct innovative designs

- TDS, incorrect innovative designs - TDS, incorrect routine designs

n TDS, correct routine designs

Fig. 7. Design sub-spaces.

Page 10: The shared expertise model for teaching interactive design assistants

620 TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

the assistant design space. The heuristic and procedural knowledge needed for a creative design is not derivable from the existing KB, even by the use of plausible reasoning.

The set union TDS U RDS t3 IDS represents the shared expertise space where both the human designer and the design assistant compose designs. The assistant proposes routine designs or innovative designs, and the designer composes creative designs. This gradation of designs helps them to exchange experience and expertise by mutual evaluations and explanations of the collaborative designs. The designer and the assistant share their expertise in the interactive design process. Composing designs from imper- fect knowledge creates opportunities to acquire new knowledge and to improve existing knowledge, as will be shown in Section 5.

4.2. The shared expertise algorithm

The design and learning processes, as well as the communication between the designer and his knowledge- based computer assistant, are controlled by the shared expertise algorithm presented in Table 5.

After receiving design specifications the assistant analy- zes them, and if some of the terms used are new, their definitions are elicited from the human designer. After the analysis phase, the design is synthesized by the assistant (in the case of a routine or an innovative design) or elicited from the expert (in the case of a creative design). During the synthesis phase the assistant may also apply various constraints and preference criteria that are part of its meta- knowledge, to choose between competing designs.

The evaluation of the proposed design is an opportunity for improving the assistant's knowledge. There are two

basic learning scenarios: learning from success (described in Section 5.3), and learning from failure (described in Sections 5.4 and 5.5).

The behavior of the assistant is dependent upon the quality and the quantity of its KB. Initially, the process of knowledge elicitation is dominant. Over time, knowledge refinement occurs more and more often. Ultimately, the assistant's KB becomes good enough so that knowledge acquisition is rarely needed. However, at any time the assistant is able to react properly to unknown input, and learn from it.

5. APPRENTICESHIP LEARNING OF DESIGN KNOWLEDGE

5.1. The learning method

To acquire new knowledge, and to refine existing knowledge, the design assistant applies a multistrategy apprenticeship learning method based on the plausible version space representation. This learning method, imple- mented in the Disciple toolkit, integrates explanation-based learning, learning by analogy, empirical inductive learning from examples, and learning by questioning the u s e r . 6'7'1°

The design assistant reacts to new input information, obtained from the human expert with the goal of extending, updating, and improving the KB to integrate the new input information. For instance, each design problem and its correct solution (indicated by the expert or generated by the agent) is regarded by the agent as a positive example of a general heuristic design rule. Conversely, incorrect solu- tions proposed by the agent will be treated as negative examples. The agent learns and modifies PVS rules based on the examples of the design episodes that it encounters

Table 5. An outline of the shared expertise algorithm

REPEAT Analyze design specifications SD

IF SD contains new terms THEN Elicit new term definitions from the designer

Synthesize design: propose and revise design extension IF SD is covered by PLB

THEN compose Routine Design DR ELSE IF SD is covered by PUB

THEN compose Innovative Design DI ELSE SD is not covered by any rule

Elidt Creative Design Dc from the designer Evaluate design

IF the design is accepted THEN Learn from Success

IF DI THEN Generalize PLB IF Dc THEN Formulate new PVS rule

ELSE Learn from Failure IF DR THEN Define exception IF DI THEN Specialize PUB

UNTIL satisfactory design .,,,i

Page 11: The shared expertise model for teaching interactive design assistants

TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL 621

and the explanations received from the designer. These modifications cause an evolution of the routine and innovative design spaces of the assistant toward the target design space. The following subsections illustrate the different learning processes invoked during design.

5.2. Learning a heuristic design rule from a new creative design

The creative design problems belong to the set difference TDS-IDS (see Figs 6 and 7) which represents the set of the designs that cannot be derived by the assistant, even by performing plausible inferences. Solutions to these prob- lems are provided by the expert, and recorded by the assistant. Each new creative design problem is an opportu- nity for the assistant to learn a new PVS rule. Adding this rule to the KB enlarges both RDS and IDS.

Assume that during the problem analysis phase the assistant needed to solve the problem shown in the top-left hand side of Fig. 8. This is the problem of proposing a design extension for a QUADRA-840av Macintosh that should have a maximum memory of 32 MB.

However, the assistant was not able to solve this problem even by applying its plausible reasoning capabilities. In such a case, the assistant treats this as a creative design problem and elicits the solution from the designer. The designer chooses to extend the description of QUADRA- 840av with SIMM-8MB-60NS-32B-72P, which is the memory module selected for the first memory bank. The problem and solution described by the designer become an initial example for a learning session.

When the expert specifies a new creative design like the one in Fig. 8, he must also explain it to the assistant. The

INPUT Problem / Situation ?P73 IS PROPDSE-DESION-EXTENSION

ARTIFACT-DESIGN ?049 MAX-MEM-LIMIT 32-MB

?Q49 IS QUADBA-B40AV OVI?~ Solution / Action

EXTEND-BY PARAMETER 7P72 ARTIFACT-DESION ?049

?P72 IS PARAMETER-DESCRIPTION MM-IN-BANKI ?S73

~S73 I~ SIMM-8MB-60NS-32B-72P

Plausible gpper Bound IF ?P73 IS PROP0 SE -DES I GN-EXTENS I 0N

ARTIFACT-DESION ?049 MAX-MEN-LIMIT ?X79

7X79 IS SOMETHING 9049 IS SOMETHING

CONTAINS ?SSI ?S73 IS SOMETHING

PLUOS-INT0 ?S81 ~MORY- C~ACITY ?XB0

?XS0 IS SOMETHINO LESS-THAN ~X79

?SSl IS S0~THIN0 Plausible Lower Bound IF

?P73 IS PROPOSE-DESI GN-EXTENS 10N ARTIFACT-DESIGN ?Q49 MAX-MEM-LIMIT ?X79

?X79 IS 32-MB 9049 IS QUADRA-840AV

CONTAINS 9S81 9S73 IS $IMM-8~-f0NS-32B-V2P

PLUOS-INTO ?S81 MEMORY-CAPAC ITY ?X80

?X80 IS 8-MB LESS-THAN ~X79

?S81 IS SIMM-SLOT-72P THEN

EXTEND-BY PARAMETER ~P72 ARTIFAOT-DESION ~Q49 IS PARAI~TER-DE S CRIPTI 0N MM-IN-BA~I ?$73

?P72

32-MB ?Q49 ?S73

* ?S73 IS S I M M - 8 1 ~ - 6 0 N S - 3 2 ~ ~ ~ 7 2 P '~'MORY-c~PAcITY 8 - ~ LESS-TH~N 32-MB, 32-MB ?S73 IS SIM~-8~-f0NS-32B-72P PRODUCT-NUMBER 11027 ~Q49 IS flUADBA-840hV BUS-STANDARD NuBu~ ?Q49 IS QUADBA-840AV CPU-CLOOK 40 ?Q49 IS flU~BA-840Ag RAM-MAX-CONFIG 128 ?049 IS 0UADRA-84[I~V RAM-MIN-CONFIO 8 7S73 IS SIMM-8~m-60MS-32B-72P I~MSRY-OAPACITY 8-~, 32-MB 0REATER-THAN 8-MB

* ?S73 IS SI~-81~B-60NS-32B-72P PLUOS-INT0 SI~-SLOT-72P, ?Q49 IS 0UADRA-840AV CONTAINS SIMM-SLOT-72P

The above e x p l a n a t i o n s have been generated

Fig. 8. Rule formulation tool.

Page 12: The shared expertise model for teaching interactive design assistants

622 TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

explanation is expressed in terms of the properties and relationships between the objects from the creative design. There are various techniques to facilitate the process of defining these explanations. For instance, the assistant can search the semantic network representing the objects from the creative design, proposing their relationships or proper- ties as plausible explanations. The expert has to choose from them the relevant ones, and can also define additional explanations using the explanation editor. Currently, there are the following types of explanations in Disciple:

• A s s o c i a t i o n - - a relationship between two objects in the example;

• C o r r e l a t i o n - - a common feature between two objects in the example;

• P r o p e r t y - - a property of an object in the example; and • R e l a t i o n s h i p - - a relationship between an object in the

example and a new object.

For the given example of creative design, several explanations are proposed by the assistant (see the bottom of Fig. 8). Two of them are accepted by the expert as relevant ones (those marked by *):

• the selected memory module SIMM-8MB-60NS-32B- 72P has capacity 8 MB which is less than 32 MB, the required maximum memory limit,

• the selected memory module SIMM-8MB-60NS-32B- 72P can be plugged into the SIMM-SLOT-72P slot with 72 pin layout, and the workstation QUADRA-840AV contains the same type of memory expansion slot.

The explanations selected by the expert are examples of the assoc ia t ion and corre la t ion explanations, respectively.

A new PVS rule is learned from the explanations proposed by the assistant or provided by the expert. ~8 The learned rule is shown in the top right-hand side of Fig. 8. The PLB of this rule is simply a reformulation of the object descriptions from the creative design, in terms of rule variables. For instance, the following component of the lower bound (see the top right pane of Fig. 8):

?Q49 IS QUADRA-840AV CONTAINS ?$81

?$81 IS SIMM-SLOT-72P

indicates that the variable ?Q49 can only take the value QUADRA-840AV, and the variable ?$81 can only take the value SIMM-SLOT-72P. Therefore, the PLB can only match the current design specification.

The PUB is an inductive generalization of the PLB obtained by turning all specific object names into generic object names, and preserving the relationships among them. For instance, the above component of the lower bound is inductively generalized to:

?Q49 IS SOMETHING CONTAINS ?$81

?$81 IS SOMETHING.

The meaning of the above condition is that the variable ?Q49 can have as value any object that is characterized as containing some other unspecified object.

The purpose of the PUB is to allow the system to propose innovative designs for specifications that are similar to the current one, in that they are both matched by this bound. Examples of such cases are presented in the next sections.

5.3. Extending the routine design space

The rule from the right-hand side of Fig. 8 is later applied to analogous designs, which are accepted or rejected by the designer. For instance, the assistant will apply this rule to the new problem specified in the bottom left-hand side of Fig. 9 because the PUB of this rule is satisfied. The corresponding design proposed is an innovative design.

The problem is to propose a design extension for another Macintosh workstation--MAC-IIVX, that should have a maximum memory of 8 MB, and contains the memory slot SIMM-SLOT-30P. The extension proposed by the assistant is a memory module for the first memory bank. This is the SIMM-4MB-80NS-8B-30P module with 4 MB capacity. This solution proposed by the assistant is correct, and is therefore accepted by the designer. The capacity of the selected module is less than the maximum memory limit, and the module can be plugged into the 30 pin SIMM slot that is part of the MAC-IIVX mother board.

In general, if an innovative design proposed by the agent is accepted by the designer, then PLB of the rule is generalized as little as possible so as to cover the current design and to remain less general than the PUB. These generalizations (indicated in the bottom part of the Fig. 9 window) are made by climbing generalization hierarchies like the one in Fig. 4. The new PVS rule, with a generalized plausible lower-bound condition, is shown in the top part of Fig. 9.

As a result of generalizing the rule's PLB, the routine design space was extended to cover more of the intersection IDS n TDS (see Fig. 7). As a result of this learning process, designs that used to be innovative are becoming routine designs.

5.4. Improving the innovative design space

Consider now another configuration problem faced by the assistant. This is the problem of proposing a design extension for a video controller board called DIAMOND- STEALTH-64-ID, as shown in the bottom left part of Fig. 10. The rule from Fig. 9 has been refined, based on learning from other configuration problems, and then has been applied to the problem shown in Fig. 10. The assistant applied this PVS rule because its PUB was satisfied. Consequently, it proposed to plug a video memory module into the video memory slot of the controller board. However, this innovative solution is rejected by the designer, who explains that the memory module selection

Page 13: The shared expertise model for teaching interactive design assistants

TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL 623

rule can be applied to computers, but not to electronic control devices such as the video controller boards.

In this case, the assistant will use the explanation provided by the expert, and will specialize the PUB of the rule to no longer cover this situation and to remain more general than the PLB. The failure explanation provided by the expert is a variable name (?Q49) representing a wrong object description in the current design episode. The part of the PUB condition describing this variable is specialized to satisfy the expert's requirement that the object for which a memory module is selected cannot be an electronic device. As a result of this specialization the value of the variable ?Q49 is constrained to be a computer. This new PUB is shown in the top part of Fig. 10.

A rejected innovative design falls into the set IDS-TDS

(see Fig. 7). The described process of specializing the PUB of a PVS rule will cause the IDS to be specialized, removing wrong designs from it.

5.5. Handling exceptions

When the assistant proposes a routine design which is rejected by the designer, the corresponding problem specifi- cation and solution description are defined in the KB as an exception. For instance, if the assistant faced a problem of selecting the type of a memory module for a PC-XT, it would apply the rule from Fig. 10 because the rule's PLB condition is satisfied (PC-XT, one of the first versions of IBM-PC-compatible computers, can be classified as a computer workstation), However, the rule would not produce a correct solution because the memory of PC-XT

RULE-PI-3-30 RULE-PI-4-2 RULE-PI-4-3 RULE-P1-4 -4 RULE-PIP-4-10 RULE-PIQ-4-5 RULE-P2-3-30 RULE-P2-4-2 RULE-P2-4- 3 RULE-P2-4-4 RULE-P20-3- 30 RULE-P22- 3-30 RULE-P3-4-3 RULE-P3-4-4 RULE -P4 -4- 3 RULE -P4 -4 -4 RULE-Pb-4-3 RULE-RI-4-10 RULE-R2-4-1@

R U L E - R 3 - , . 4 - 1 0 RULE-R4D-4-10

l ..... ~.~::',: :;: :::~..~ :::::::::::::::::::::::::::::::::::::::::::::::: INPUT Problem / Situation

? P 7 3 I S AKTIFACT-DESISN MAX-MEM-LINIT

?Q49 IS coNTAINS

?S81 IS OUTPUT Solution / Action

EXTEND-BY PAF~hNETER ARTIPACT-DESI GN

~P72 IS MM- IN-BANKI

?$73 IS PLUGS-I}Ff0 MEMORY- CAPACITY

4 -I~B L E S S - T H A N

Plausible Upper Bound IF ?P73 PROPOSE-DESION-EXTENSION . ARTIFACT-DESIGN ?Q49 , MAX-MEM-LIMIT ~X79 ~X79 SOMETHIN0 ?Q49 $0METHIN0 , CONTAINS ?S81 ?573 soMETHING , PLUOS-INTO ?S81 . MEMOBY-CAPACITY ~XSO ~X80 SOMETHINO , LESS-THAN ?X79 ?S81 SOMETHIN0

Plausible Lower Bound IF ?P73 PROP0SE-DESIGN-EXTENSION , AKTIFACT-DESIGN ?Q49 . MAX-M~M-LIMIT ?X79 ?X79 MB-CAPACITY ?Q49 MACINTOSH-WOBKSTATION , CONTAINS ?581 ?$73 RAM-SIMM , PLUOS-INTO ?S81 , MENORY-CAPACITY ~X80 ?X80 NB-CAPACITY , LESS-THAN ?X79 ?S81 SIMM-SLOT

THEN EXTEND-BY PARAMETER ?P72 , ARTIFACT-DESION ?049 ?P72 PARAMETER-DESCRIPTION , MM-IN-BANKI ?$73

PROPOSE-DES I ~N-EXTENSION 7049 8-MB MAC-IIVX ?$81 SIMM- $LOT-30P

7P72 ?O49 PARAI~TER-DE S CRIPTI 0N ?S73 S IMM-4MB- 80NS- 8B - 30P ?$81 4-MB 8-MB

INPUT Problem / Situation ?P73 IS PROPOSE-DESION-EXTENSION

ARTIPACT-DESION 7049 MAX-MEM-LIMIT 32-MB

?049 IS QUADKA-O40AV CONTAINS ?S81

?S81 IS 5D~-SLOT-72P oUTPUT Solution / Action

E~"TEHD-BY PARAMETER ?P?2 ARTIFACT-DESIGN ?049

?P72 IS PARAMETER-DESCRIPTION MM-IN-BANKI 9$73

?S73 IS $IMM-8MB-60NS-32B-72P PLUGS-INT0 7S81 ~M0~Y-CAPACITY 8-MB

8-MB LESS-THAN 32-MB

~I made the f o l l o w i n g g e n e r a l i z a t i o n ( s ) : '~32-MB 8-MB - -> M B - C ~ A O I T Y QUADRA- 840AV MAC- II%'X -- > MACINTOSH-WORKSTATION SIMM-8MB-60NS-32B-72P SIMM-4MB-80N$-8B-30P - - > RAM-SIMM

[MI 9-6-c

F i g . 9 . Rule refinement tool--accepted example.

Page 14: The shared expertise model for teaching interactive design assistants

624 TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

was not configured from memory modules. Instead, it used to be configured from memory chips plugged directly onto the mother board. In such a case, the PLB condition cannot be specialized to uncover this negative example because such a specialization will cause the uncovering of some positive examples of the rule.

The described exception falls into the set difference RDS- TDS (see Fig. 7). This is the set of designs that are deductively derivable from the KB, but are incorrect. This shows that there are errors in the set of facts and deductive rules in the KB which will generate incorrect solutions to routine problems. This is one of the most dangerous situations in the assistant's behavior, in which deductive inferences produce incorrect results.

Such covered negative examples point to the incomplete- ness of the assistant's knowledge, and are used to guide the elicitation of new concepts and features, by using the knowledge-elicitation methods described in Ref. 7. The exception handler of the Disciple Toolkit is a special tool that deals with this kind of case.

6. CONCLUSIONS

This paper has presented the shared expertise model for interactive problem solving and learning in the context of engineering design. This model is at the basis of building learning assistants with the Disciple toolkit that has also been briefly presented in this paper. The Disciple toolkit is

RULE-PI-3-30 RI/LE-PI-4 -2 RULE -P 1-4 - 3 RULE-P1-4-4 RULE-PIF-4-10 RULE-P18-4-5 RULE-P2-3-30 RULE-P2-4-2 RULE-P2-4-3 RULE-P2-4-4 RULE-P20-3-30 RULE-P22- 3-30 RULE-P 3-4- 3 RULE-P3-4-4 RULE - P4 - 4 - 3 RULE -P4 -4 -4 RULE-P5 -4- 3 RULE-RI-4-1D RULE-R2-4- i0

RULE-R3-4-10

Plausible Upper Bound IF ?P73 PROPOSE-DESIGN-EXTENSION , ARTIFACT-DESIGN ?Q49 , MAX-MEM-LIMIT ~X79

?X79 ME- CAPACITY ?849 COMPUTER , CONTAINS 7S81 ?S73 MEMORY-MODULE , PLUSS-INTO ?S81 , MEMORY-CAPACITY ~X80 ?X80 SOMETHING , LESS-THAN 9X79 ?$81 SOMETHIN8

Plausible Lower Bound IF ?PT~ PROPOSE-DESIGN-EXTENSION , ARTIFACT-DESIGN 9049 , MAX-MEM-LIMIT 7X79

",X79 MB- CAPACITY ?Q49 MACINTOSH-WORKSTATION . CONTAINS 7S81 ?$73 RAM-MODULE , PLUGS-INT0 ?$81 , MEMORY-CAPACITY ~XS0 ?X80 MB-CAPACITY , LESS-THAN ~,X79 ?581 RAM-SLOT

THEN EXTEND-BY PARAMETER ?P72 , ARTIFACT-DESIGN ?049 ?P72 PARAMETER-DESCRIPTION , MM-IN-BANKI ?$73

INPUT Problem / Situation II@UT P~oble~ / Situation ?P73 IS PROPOSE-DESIGN-Ek"IJ£NSION ?P73 IS

ARTIFACT-DESIGN ?049 ARTIFACT-DESION MAX-MEM-LIMIT 730-MB MAX-MEM-LINIT

7049 IS DIAMOND-STEALTH-64-1D 7049 IS coNTAINS ?$81 cONTAINS

7S81 I5 VIDE0-DIMM-SLOT ?$81 IS OUTi~tIT Solutlon / Action 0LrI~UT Solutlon / Actlon

Ek"YEND-BY P~R~TER 9p72 EXTENB-BY P ~ T E R ARTIFACT-DESIGN ?849 ARTIFACT-DESIGN

?P72 IS pAR~T£R-DESCRIPTION ~P72 IS ~-IN-BANKI ?$73 MM-IN-BANKI

?$73 IS VBIMM-1000K-70NS ~$73 IS PLUSS-INTO ?$81 PLUGS-INT0 MEMORY-CAPACITY I-MB

I-MB LESS-THAN 720-MB 8-MB

PROPOSE-DESIGN-EXTENSION 7049 32-~ QUADRA-840AV ?581 SI~IM-SLOT-72P

?P72 7049 PARAMETER-DES CRIPTI 0N ?$73 5INM- 8MB-60NS-32B- 72P ?$81

M~MOKY- CAPAC ITY 8-14B LESS-THAN 32-MB

I made the followlng particularization(s) for not covering DIAMOND-STEALTH-64-1D

SOME~'dING -- > COI~I2/ER

Fig. 10. Rule re f inement t o o l - - r e j e c t e d example .

Page 15: The shared expertise model for teaching interactive design assistants

TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL 625

being implemented using the Harlequin LISP software engineering environment on SUN platforms. Most o f the tools are already functioning, and they have been tested and used to build various agents. The toolkit is an extension and a modularization o f the authors' earlier systems, Disciple and NeoDisciple, which were implemented in Macintosh LISP.

In the shared expertise model, domain-specific problems are solved cooperatively by a human expert and his KB design assistant. This cooperation between a human designer and his design assistant increases the efficiency of the design process.

An important strength of SEM is that it takes advantage o f a key feature that typically characterizes a human e x p e r t ~ e can easily solve problems in his area o f expertise, but has difficulties in expressing his knowledge as general problem-solving rules. Indeed, what is required from the expert is only to give examples o f correct designs, and explanations o f these examples, and to judge if designs proposed by the assistant are correct or not. From such interactions the design assistant is able to learn general heuristic design rules, and to improve its KB. Moreover, this type o f knowledge acquisition and learning, performed during the interactive design process, allows the acquisition process to be focused on acquiring only the needed knowledge.

The shared expertise model also has several weaknesses. The design process is divided into several phases in order to evaluate composed designs and to identify situations when the knowledge-acquisition and learning modules should be invoked. This may significantly slow down the design process, because the expert may need to teach his assistant. Also, some of the results of the design problem-solving process should be evaluated by the assistant itself, instead of relying entirely on the designer. This, however, requires a better definition o f the assistant's model for handling uncertainty. These are very important topics for future research.

The problem of storing and maintaining the problem- solving episodes as learning examples needs to be addressed more carefully. The current solution allows for the recording of the expert's problem-solving steps, but it does not provide the opportunity to change already recorded steps. It is generally agreed that experts change their minds during the problem-solving process, and they should have the opportu- nity to reflect these change when they cooperate with the design assistant.

Acknowledgements--This research was partially supported by the Advanced Research Projects Agency Contract No, N66001-95-D-8653 and the NRC Twinning Program with Romania.

REFERENCES

1. Mitchell, T. M., Mahadevan, S. and Steinberg, L. 1., LEAP: a learning apprentice system for VLSI DESIGN. In Proceedings of the Inter- national Joint Conference on Artificial Intelligence (IJCAI.85), Los Angeles, 1985. Morgan Kaufmann, San Mateo, CA, pp. 573--580.

2. Tecuci, G., Disciple: a theory, methodology, and system for learning expert knowledge. Ph.D. thesis, University of Paris--South, 1988.

3. Wilkins, D. C., Knowledge base refinement as improving an incorrect and incomplete domain theory. In Machine Learning: An Artificial Intelligence Approach, Vol. Ill, ed. Y. Kodratoff and R. S. Michalski. Morgan Kaufmann, San Mateo, CA, 1990, pp. 493--513.

4. Michalski, R. S. and Tecuci, G. (Eds), Machine Learning: A Multistrategy Approach, Vol. IV. Morgan Kaufmann, San Mateo, CA, 1994.

5. Cypher, A. (Ed.), Watch What I Do: Programming by Demonstration. MIT Press, Cambridge, MA, 1993.

6. Tecuci, G., Automating knowledge acquisition as extending, updating, and improving a knowledge base. IEEE Trans. on Systems, Man and Cybernetics, 1992, 22(6), 1A.~4 1460.

7. Tecuci, G. and Hieb, M., Consistency-driven knowledge elicitation: using a learning-oriented knowledge representation that supports knowledge elicitation in NeoDisciple. Knowledge Acquisition Journal, 1994, 6(1), 23--46.

8. Tecuci, G. and Duff, D., A framework for knowledge base refinement through multistrategy learning and knowledge acquisition. Knowledge Acquisition Journal, 1994, 6(2), 137-161.

9. Tecuci, G. and Kodratoff, Y. (Eds), Machine Learning and Knowledge Acquisition: Integrated Approaches. Academic Press, New York, 1995.

10. Tecuci, G. and Kodratoff, Y., Apprenticeship learning in imperfect domain theories. In Machine Learning: An Artificial Intelligence Approach, Vol. 3, ed. Y. Kodratoff and R. S. Michalski. Morgan Kaufmann, San Mateo, CA, 1990, pp. 514-551.

11. Dybala, T. and Tecuci, G., Shared expertise space: a learning-oriented model for computer aided engineering design. International Joint Conference on Artificial Intelligence (IJCAI-95), Machine Learning in Engineering Workshop, Montreal, Canada, 1995.

12. Dybala, T., Shared expertise model for building interactive learning agents. Ph.D. dissertation, George Mason University, Fairfax, VA, 1996.

13. Coyne, R. D., Rosenman, M. A., Radford, A. D., Balachandran, M. and Gero, J. S., Knowledge-Based Design Systems. Addison-Wesley, Reading, MA, 1990.

14. Gero, J. S., Design prototypes: a knowledge representation schema for design. AI Magazine, 1990, 11(4), 26-36.

15. McDermott, J., RI: a rule-based configurer of computer systems. Artificial Intelligence, 1982, 19(1), 39-88.

16. Marcus, S. and McDermott, J., SALT: a knowledge acquisition language for propose and revise systems. Artificial Intelligence, 1989, 39, 1-37.

17. Tecuci, G., Cooperation in knowledge base refinement. In Machine Learning: Proceedings of the Ninth International Conference (ML92), ed. D. Sleeman and P. Edwards. Morgan Kaufmann, San Mateo, CA, 1992, pp. 445-450.

18. Tecuci, G., Hieb, M. R. and Dybala, T., Teaching an automated agent to monitor the electrical power system of an orbital satellite. Telematics and lnformatics Journal, 12, 229-245.

AUTHORS' BIOGRAPHIES

Tomasz Dybala is a researcher at the Computer Science Department of George Mason University, U.S.A. He has finished his Ph.D. thesis, under the direction of Professor Tecuci. He has published a dozen technical papers in the areas of machine learning, applications of knowledge-based technologies to engineering design, and control engineering. He has been involved as a researcher and as a system developer in several R&D projects, both in the university and in industrial environments. He is currently working on the development of the Disciple toolkit and its applications to the building of personal design assistants and intelligent tutors for software packages. Gheorghe Tecuci is Professor of Computer Science at George Mason University, and a member of the Romanian Academy. He has published over 85 scientific papers, and contributed to the development of two new research areas in AI: multistrategy learning, and integrated knowledge acquisition and machine learning. He co-edited two books: Machine Learning: A Multistrategy Approach (Morgan

Page 16: The shared expertise model for teaching interactive design assistants

626 TOMASZ DYBALA et al.: SHARED EXPERTISE MODEL

Kaufmann, 1994), and Machine Learning and Knowledge Acquisition: Integrated Approaches (Academic Press, 1995). He was a co- organizer of the first workshops on these new research directions in AI, and presented the first tutorials on these topics at the IJCAI and AAAI conferences. I-Iadi D. Rezazad is a Ph.D. candidate in Information Technology at George Mason University. He is conducting his research under the direction of Professor Tecuci. He received his master's degree in Computer Science, and his bachelor's degree in Computer Science, Statistics, and Applied Mathematics from The American University. He is the Chief Executive Officer of Orchid Technologies and Management, L.C., a Washington, D.C. based computer consulting company which specializes in systems integration, systems design and development, and application programming. His current research is in machine learning, knowledge acquisition, and intelligent learning agents, and their engineering applications.