ASME 2000 Design Engineering Technical Conferences

Proceedings of DETC’00

ASME 2000 Design Engineering Technical Conferences

and Computers and Information in Engineering Conference

Baltimore, Maryland, September 10-13, 2000

DETC2000/CIE-14617

1Copyright © 2000 by ASME

Signposting: an AI Approach to

Supporting Human decision Making in design

1Copyright © 2000 by ASME

Abstract

Artificial intelligence provides powerful techniques for formalising the art of engineering problem solving: for modelling products, describing task structures, and representing problem solving expertise as inference knowledge and control knowledge. Signposting systems extend the scope of these methods beyond automatic design by using them to provide both information and guidance for decision-making by human designers. This paper outlines the application of AI methods according to cognitive engineering considerations, to the development of knowledge management tools for engineering design. These tools go beyond conventional knowledge management and decision support approaches by supplying both inference knowledge and strategic problem solving knowledge to the user, as well as information about the state of the design. By focusing on tasks and on the dependencies between design parameters, signposting systems support contingent and flexible organisation of activities. Such tools can support product modelling, design process planning and capturing expert design knowledge, in a form that can be used directly to guide the organisation of design activities and the performance of individual tasks. A key element of this approach is the incremental acquisition of product models, task structures and problem solving knowledge by defining variant cases.

1Introduction

One way to view symbolic artificial intelligence is that it is the attempt to understand the art of problem solving in rigorous computational terms. This view of the AI project, focusing on knowledge and mechanism, is broader than the behaviour-centred view of AI as the construction of intelligent systems. It offers a different perspective on how to apply the concepts and methods of AI to problems too subtle and open-ended for independent AI reasoning systems, such as conceptual design in engineering, which we are applying in the development of signposting systems.

The demands and pressures on many engineering designers are rapidly increasing, as they have to obey more and more regulations, consider a wider variety of factors and best practice procedures (see for instance Huang, 1996), and strive ever harder to reduce lead times and avoid mistakes that cause costly revisions. But many engineering companies possess a great deal of experience and expertise; the challenge is to retain this expertise despite staff turnover, develop it, and deploy it where and when it is needed. But how can effective corporate knowledge management in design be achieved? While expert designers have a wide knowledge of facts and examples, their essential expertise lies in skills for analysing and solving particular kinds of problems. While obtaining information is a major drain on engineers’ time, possessing the necessary range of expertise to integrate this information considering all the important issues can be a much less tractable problem. Knowledge management and decision support systems can provide far more assistance to designers by supporting problem analysis and design synthesis as well as fact gathering.

AI has been applied to corporate management of problem solving expertise: one important commercial purpose for expert systems is to make explicit and encode the problem-solving skills of human experts for reuse by non-experts. This can be motivated by the desire to preserve and exploit the intellectual assets of a company whose employees may leave or retire.

Automatic design systems can now tackle some significant tasks, but in many situations where companies might want to encode and reuse the skills of expert designers, using automatic design systems is impossible or yields inadequate results. The skills of expert designers include the perceptual recognition and evaluation of subtle features that defy analysis and computational representation (see Eckert et al., 1999a). But these perceptual skills are combined with knowledge of what inferences to make and what procedures to follow that can be articulated, recorded and passed on.

Figure 1 The Signposting concept

The approach (illustrated in Figure 1) taken to capturing and reusing expert problem solving knowledge in signposting systems is to combine knowledge bases created with the knowledge acquisition and knowledge description techniques of artificial intelligence, with an inference engine capable of complex pattern analysis and synthesis operations, and solving ill-structured problems with incomplete information: a human designer. Signposting systems provide a human user with problem-state information, the task structure and the state of the design. They also provide within-task problem-solving knowledge in the form of design guidelines, and strategic problem-solving knowledge in the form of task selections and rankings, and task selection guidelines. They thus make corporate knowledge of how to design a resource for designers to draw on, actively presented when apposite to the problem in hand.

This approach applies to decision support the analyses of the structure of problem solving developed for knowledge engineering for expert systems, notably in the CommonKADS methodology (for instance Schreiber et al., 1993, 1999), emphasising the distinctions between domain knowledge, inference knowledge, task knowledge and strategic knowledge. As Smithers (1996, 1998) argues, modelling design activities in terms of the knowledge and information agents require to perform them provides the type of understanding required for building effective design support systems.

Section 2 introduces the signposting idea in the context of its antecedents. Section 3 discusses how signposting can be used to manage expert design knowledge, and points out some cognitive engineering issues that have to be addressed in developing design support systems. In section 4 we outline the architecture of a signposting system we are currently implementing, that takes an AI approach to knowledge management for design; we discuss in passing how the signposting approach relates to problems stemming from the limitations of human cognitive abilities and organisational practices.

2Signposting design tasks

The signposting approach to supporting design decision making comes out of the realisation that many important design processes have structures that defy conventional linear process descriptions. They involve complex interdependencies between design choices, so that designers have to estimate parameter values, backtrack, and repeat some tasks many times before all the parameters have satisfactory and mutually consistent values, even when what the parameters are is well understood. Such processes can be complex and important enough to require systematic planning, and challenging enough to gain from expert guidance and decision support.

The core ideas of signposting (Clarkson and Hamilton, 1999, in press) emerged during a study of a good example of such a process: helicopter rotor design (Hamilton, 1999; Hamilton et al., 1997). Form, material and production method are indivisible in a rotor blade, which may be used for years in diverse environments. About 40 engineers with a wide range of expertise participated in the rotor blade development project observed by Hamilton, and no one had a good understanding of all aspects of the design, despite the fact that the company is a world leader in rotor blade design. This made project management and planning extremely difficult. While the fundamental stages of the design process were easy to identify, more detailed analysis revealed a nest of information dependency loops and intertwined and repeated tasks. Process-centred modelling techniques such as those presented by Pahl and Beitz (1996) proved too coarse-grained or failed to capture the nature of the interdependencies.

Signposting is founded on the hypothesis that in situations involving complex interdependencies between different aspects of a design, both general-purpose problem structuring techniques and domain-specific expert knowledge can be applied to achieve a linear ordering of tasks and design decisions. This involves creating sequences of tasks to make collections of interdependent decisions, by making estimates of parameter values and then using those estimates to refine each other until they converge to a set of satisfactory and mutually consistent values. Signposting embodies the further hypotheses that designers and managers could benefit from timely advice from a computer support system both about identifying and choosing tasks, and about how to perform them. This guidance is supplied by recognising situations when tasks are possible, and when guidelines are appropriate.

This assumes that one can identify the design decisions that need to be made, that is, the parameters of the design. The signposting approach is being applied initially to variant design tasks, where the form of the product is known from previous designs, but obtaining the right parameter values is highly complex. We are now generalising the approach to cover the design of ranges of similar products, where identifying the parameters of a design and their interactions is an integral part of designing (see section 3). In addition we are investigating customisation and change design, again important in the aerospace industry, where creating new products involves patching or partial redesign, and estimating the scope of the changes required to achieve a modification is essential (Eckert and Clarkson, 2000; Clarkson et al., 2000b).

2.1Units of analysis: parameters, dependencies and tasks

Key concepts in the signposting approach are tasks, parameters, dependencies and confidences, for which we use the operational definitions described here.

A parameter is an aspect of a design that needs to be determined, and hence embodies a decision about the design. While parameters may have single numerical values, we don’t limit the term to such simple cases. Parameter values may be symbolic, or have internal structure (such as complex shapes), or be clusters of related parameters that we want to treat as a unit. Users can model products simply as flat parameter lists, but the more advanced signposting system proposed in section 4 is designed to support hierarchical product modelling. We treat derived information about a design, that does not embody any decisions, as derived parameters; whether or not they are included in product models or in task descriptions is an issue of convenience.

The values of parameters are often dependent on other parameters. A dependency is a causal or constraint relationship linking one parameter to another. Dependencies can be one-way or mutual; we treat mutual dependencies as pairs of one-way dependencies.

The confidence in a parameter value is an indication of how far it is to be trusted as an indication of the final value. How to describe confidence in a form both suitable for AI uncertainty reasoning and intuitive and useful for designers is a subject of ongoing research (see section 4.4). The current running signposting systems use simple unitary qualitative confidence values.

Tasks are the units of analysis of design processes. A task is an activity that takes the values and confidences of certain parameters as inputs, and generates or updates the values and confidences of other parameters (possibly including the input parameters). Task descriptions include confidence matrices (see Figure 2), naming their input and output parameters, and describing the confidences required for the input values and expected for the output values. We envisage that the users of the advanced signposting system outlined in section 4 will describe tasks hierarchically, first creating high level tasks and then adding the subtasks that are performed within them.

Figure 2 Confidence Matrix

Signposting systems support decision making within tasks by presenting guidelines. Guidelines have no fixed form; but they are typically descriptions of principles, considerations, and problem-solving procedures that are useful for performing the currently active task.

2.2Designing with interacting parameters

In many important design processes, performing variant design and change design, much of the structure of the design is known in advance. Moreover, a lot of conceptual design proceeds middle-out: major elements are selected from available products, or imagined in considerable detail based on prior experiences of similar systems, and the rest of the design is created around them. In these situations, the parameters of the design are known or can be identified relatively straightforwardly. What can make such processes difficult and complex are the interactions between the parameters and tight, conflicting constraints on the form of the design. Conceptual-onwards process analyses, for instance following Pahl and Beitz (1996), and the design support systems based on them, for instance PROSUS (Blessing, 1994) are too coarse-grained for unpacking the activities involved in finding consistent parameter values. Similarly abstract-downwards approaches, for instance from Andreasen (1980), and design support systems using functional models, for instance Schemebuilder (Bracewell and Sharpe, 1996) and MAX (de Vries, 1994), are irrelevant to these problems.

By using tasks as the unit of analysis, design processes can be modelled bottom-up from known atomic tasks, or middle-out from larger tasks, where considering input-output relationships and decomposing them into subtasks reveals holes in the analysis. This approach enables the analyst to recognise situations where the same goal can be achieved by alternative task sequences, and where the same generic tasks are performed repeatedly and in different situations, for instance finite element analysis or stress calculation. Modelling tasks in terms of their input-output behaviour enables them to be selected according to the states of their input parameters. In the prototype signposting system (Clarkson and Hamilton, 1999, in press) a task is highlighted as a productive next step if (a) the users have sufficient confidence in the input parameters, and (b) if the task will increase confidence in one or more output parameters. In the interface shown in Figure 3, the traffic light colours red, amber and green indicate the progress that can be made by peforming a task: red means ‘insufficient information to do the task’; amber means ‘possible, but would not advance the design’; and green means ‘possible and useful’.

Figure 3 Signposting interface

Designers do not always need to have high confidence in the accuracy and reliability of the information they use; they often make provisional decisions based on partial information, and the results of these decisions are used to reconsider the decisions and assumptions they were based on. Clarkson et al. (2000a) discuss an example of the prototype signposting system supporting this refinement process. Novice designers using the tool were better able to reach a design solution than a control group without it, and followed a task sequence closer to that of an expert.

The dependencies between the parameters are used to identify the linear order of tasks and decisions. But when parameters are directly or indirectly interdependent, the task ordering contains loops (see

Figure 4). Here heuristics for breaking dependency loops, plus domain-specific expert knowledge, are needed to propose tasks for estimating parameter values and progressively refining them until consistent and satisfactory values are established.

Figure 4 Dependency relationships between tasks

In complex real-life design processes, many tasks may be competing for limited resources, and the order in which parameter values become available cannot always be predicted. Some tasks may be delayed, or prove harder than expected, so parameter values may not be generated at the expected times with the expected confidences. The design process has to be adapted on the fly. The signposting systems encourage designers to perform the most urgent tasks that are possible with the information available.

2.3Related approaches to supporting engineering design

A variety of approaches to computer support for design have used process models comprising networks of tasks. Traditional workflow systems provide a framework for controlling business processes and are used to mediate the flow of responsibility in those processes from person to person and from task to task (Prasad et al., 1998). This matching of resource to need is suited to well behaved business processes where a standardisation of procedure can bring about increases in process efficiency. However, the engineering design process is not a well behaved process and even in the case of variant design, where a new artefact is very similar to previously designed products, there can be considerable changes in the process that generates the design information (Dong and Goh, 1998). Signposting provides an approach to manage such processes by enabling a situation-driven guidance engine.

As well as being used for retrieving cases for adaptive design using case-based reasoning (Göker, 1999), networks of dependencies between parameters have been used for computing task networks for activity planning.

Petri net models. McMahon and Xianyi (1996) use petri nets (essentially, directed graphs whose nodes are functions and arcs are parameter values) to create parameter driven design process models. In a parameter driven process the design process defines task sequences which are executed in response to parameters being available. Contextual task knowledge is not required (or at least, not included in the model). However, such task network models are static in nature and must be carefully defined for each new type of product.

Design Structure Matrices. DSMs (Steward, 1981), represent information dependency relationships as cells in matrices whose rows and columns are the independent and dependent parameters; in this notation the functions that are nodes in Petri nets are only implicit. They have been used by Eppinger (for instance Eppinger et al., 1994) and others to generate task networks as prescriptive process models for management purposes; this involves explicitly defining the precedence order of tasks. The signposting approach extends this by introducing the notion of parameter confidence as a means to differentiate between similar tasks and break dependency loops through estimating and iterative refinement. The dynamic nature of task selection in signposting does away with the need to explicitly define task precedence. Signposting can also include the use of requirements and other contextual information in task selection.