Cognitive Tools for Design Engineers: A Framework for the Development of Intelligent CAD Systems

1 Introduction

When conceptualizing design engineering as a creative problem solving process, the question arises as to the underlying cognition, specifically, how information processing of the design engineer can be supported during the development of solutions. Can we build solutions ontologies / libraries that capture the reasoning behind designs and use them to support the problem solving process of design engineers? Can they be implemented in databases that learn from users and that self-populate with user data (user approaches to their solutions)? These ontologies have a strong dynamic component because knowledge grows and reasoning can change over time. How can the temporal component be captured and implemented? How can dynamic ontologies be used to make inferences about user schemas? Can the temporally derived schemas (i. e., problem solving patterns) be used to model the continuum or majors shifts in the novice / expert continuum? Can we determine schema differences, overlap, absence or presence and use them to interpret user expertise level?

The problem solving support tools in mechanical engineering are typically traditional drawing and sketching, and over the last 10 years, CAD (computer assisted design) software, has become the primary tool of design engineers to develop and articulate their solutions. CAD tools are sophisticated and popular as computerized drafting tools but questions remain unanswered whether CAD applications support the cognitive aspects of problem solving. As early as 1989 (Wood 1996) observed that CAD tools provide computerized versions of traditional drafting and that these approaches had not yet realized the possibility of supporting the design process beyond the mechanics of drawing. Furthermore, (Wood 1996) observed that CAD applications did not support problem solving and solutions finding, specifically in the context of function based thinking, which is instrumental to design engineering.

In this paper we develop the cognitive psychological background, a rationale for CAD enhancement with cognitive tools and the research framework for cognitive CAD tools that support creative problem solving processes through reasoning and meaningful design alternatives. We present a research framework that is extensible to design engineering in general. At the core of the research plan are the development and implementation of an artificial memory that is interpreted with real-time data analyses supported by machine learning, and made accessible to the design engineer through interaction design for intelligent CAD (iCAD). The paper is structured as follows: Part 1 introduces a selection of the cognitive phenomena and biases that affect information processing during design engineering; Part 2 presents an overview of the engineering design process, including a subsection on why reasoning matters during design and the state of CAD software based on Parametric Technology’s Creo 3.0. Part 3 presents the transdisciplinary research framework to guide the development of a research plan for improving CAD with cognitive tools.

2 Part 1: Cogntive Phenomena and Biases

Well-studied phenomena occur during ill-defined problem solving. One of these is a fixation of the typical use of an object. For example, a hammer is made for hammering, a toothbrush for cleaning teeth, etc. Karl Duncker discovered the phenomenon that people tend to limit the uses of previously encountered objects in 1945 and termed his observation Functional Fixedness (Duncker 1945).

The phenomenon of functional fixedness persists as an artifact of human information processing. For example a knife can easily be used as a screw driver. The realization that the function of a knife is not limited to preconceived notions but that its uses depends on the characteristics of the object, is not necessarily obvious to the person who needs a screw driver but only has a knife at hand.

Duncker’s original study investigated functional fixedness using ill-defined problems that required productive as opposed to reproductive, or algorithmic thinking (Wertheimer 1959). The critical observation of the research is that problem solvers implicitly place limits on the functions of an object that are related to its current use, context and prior knowledge. This is related to schema use in cognition where a previously learnt framework guides recall based on a set of cues (Bartlett 1932, Marshall 2007). In this context cues are the current function or usage of an object which appears to inhibit the exploration of alternative uses, i. e., functions or functionalities. By making alternative uses of the object more obvious, problem solvers are more likely to find solutions (Wertheimer 1959).

Another phenomenon that is related to the obstacles of ill-defined problems, is analogical problem solving. Analogical problem solving refers to the transfer of a solution from one problem to another problem that seems superficially unrelated. Studies by (Gick and Holyoak 1980) showed that without specific hints participants have difficulty noticing similarities between problems. (Holyoak 1990) suggests that failure to recognize the commonalities between the problems is the result of deep vs. surface structure processing (Chomsky 1969), where surface refers to appearance and deep refers to an abstraction of structural organization. Consequently, seemingly unrelated components can prevent participants from shifting their focus from the surface to the common, deep structure of the problem unless explicitly instructed to do so. The failure mechanism is simple: dissimilar domains create different appearances (surface structures) and these differences imply that the problems have nothing in common, hence analogical search is not conducted. Appearances, whether they take the form of a problem context or the current usage of an object, lead problem solving to adopt a particular mindset (Einstellung) (Luchins 1942).

Another phenomenon similar to Einstellung and Functional Fixedness, is solution fixation, which is a common cognitive artifact affecting design engineers from the novice to the expert. Prior research (Ullman, Stauffer, and Dietterich 1987, Ullman, Dietterich and Stauffer 1988) focused on professional mechanical design engineers and found that they became fixated upon preliminary solution ideas and failed to consider alternative design concepts. The phenomenon was seen both at the level of the overall design problem and at the level of each individual sub-problem. In addition, (Ullman, Stauffer, and Dietterich 1987, Ullman, Dietterich and Stauffer 1988) observed that if the designer discovered weaknesses in original design later in the process, they were solved by ‘patching’ the design rather than discarding the idea and developing a new concept. According to (Ullman, Stauffer, and Dietterich 1987, page 15): “The first idea was almost sacred, and sometimes even highly implausible patches would be applied to make it work.” Similarly, (Ball, Evans, and Dennis 1994) in a study of pre-expert electronics designers observed that individuals rarely generated and modelled alternative solutions but focused upon initial ideas that were iteratively improved until they reached a state that was adequate. This perseverance on a specific solution path may be related to another phenomenon observed in decision making, called sunk cost (Kahneman and Tversky 1979, Arkes and Blumer 1985). Sunk cost refers to the tendency to continue with an endeavor because an initial investment has been made that is not recoverable. Similarly, design engineers invest critical resources, time and cognitive effort, into the initial solution and changing to an alternative solution creates the perception of having wasted time and effort.

In summary, design engineers “patch” and “repair” solution ideas instead of questioning or exploring alternative uses of the parts and assemblies they created. This approach and mindset during the design process limit the functionality of the design elements to the originally conceived design context: in other words, the design engineer becomes function and solution fixated. Problem context and the current object usage create cognitive artifacts such as surface structure processing, solution fixation and functional fixedness.

The phenomena presented in this section have implications for the presentation and organization of CAD tools and the way they are presented to the user.

3 Part 2: Prior Research and CAD Tools for The Engineering Design Process

The engineering design process is the mechanism that incorporates creativity, the scientific method, mathematics, physics and engineering principles to solve a particular problem. The design process is usually represented as either a design cycle or as a design spiral, where each step moves the concept forward or reverts to a previous stage before further design steps are made. The basic cycle or spiral is 1) Problem Definition, 2) Brainstorming, 3) Preliminary Design, 4) Design and Analysis, 5) Design Refinement, and 6) Design Implementation. Where steps 1 through 5 are often grouped as “Conceptual Design,” where the designer is primarily concerned with initial customer and engineering constraints and the specification of form; “Layout Design / Embodiment,” where the focus of the design changes to specifying the components (i. e., individual parts and assemblies) at decreasing levels of abstraction; “Detail Design,” where the design becomes finalized; and “Catalog Design,” where features, parts, or assemblies are selected from a library database (Note: catalogs play a significant design role by allowing the designer to select design possibilities with respect to functionality). These are combined to various degrees in current day Computer-Aided Design and Manufacturing software.

In the 1980’s and early 90’s Computer-Aided Drafting (CAD) was the primary use for computers in the development of a component (i. e., during Design Implementation). With the coming of Parametric Technology’s Pro / Engineer in the early 1990’s Computer-Aided Design and Development (CADD) was developed, where the software was integrated into the Preliminary Design, Design and Analysis, Design Refinement and the Design Implementation. This was later renamed to CAD / CAM as manufacturing was included into the design process. Parametric Technology Corporation (PTC) has gone through many iterations of Creo; its latest version Creo 3.0 is one of the most widely used CAD / CAM software programs in the world today. It is based on parametric design principles, which require that design operations are captured and stored as they take place and that any committed changes are propagated throughout the design. As such PTC Creo 3.0 maintains a history-of-changes that the designer makes in order to keep track of operations that depends on each other, but not as to “why.” Although the ability to render the design and follow parametric deign principles are valuable characteristic of any design tool, it seems that the cognitively supportive system that must include information on the deign intent that underlies a solution or a step in the solutions, the why.

4 Part 3: What can we do to Improve CAD with Cognitive Tools Using Knowledge Ontologies?

The enhancement of any CAD tool with reasoning support is not a trivial task. It requires the management of cognitive phenomena and biases during problem solving, the incorporation of reasoning support and the use of relational systems for knowledge ontologies. To investigate the emerging possibilities for cognitive tools, a number of research challenges must be met before CAD systems can support and enhance the problem solving and creative processes of design engineers.

Our initial focus of this work was on the design processes in mechanical engineering but during the development of the research agenda it became apparent that reasoning support is applicable to the majority of design engineering, because the underlying cognitive processes are shared across design disciplines. Hence the three research challenges presented next serve as a framework for any application domain and act as the transdisciplinary schema for development of a research agenda.

The three challenges in the development of a reasoning support system that enhances the problem solving process of design engineers are:

1. Building an Artificial Memory that contains reasoning ontologies

   1. Collection of Semantic Data,

   2. Temporal Data Modelling and Storage,

2. Dynamic Data Analyses and Machine learning, and

3. Interaction design for intelligent CAD (iCAD).

The first challenge is the development and construction of an artificial reasoning memory. The first step under this challenge is the collection of semantic data, i. e., the rationale for a design by the designer and the justifications and intentions that motivate specific design decisions. Some of the research opportunities to collect these data are:

1. Integration of Design processes with explicit measures of reasoning. Data can be harvested from CAD users, capturing reasoning data using verbalization protocols during the design process (self-narration) and after the design process by reviewing the steps in the product design and documenting the explanations using context reinstatement interview techniques (Bahr and Ford 2011).

2. Integration of Design processes with implicit measures of reasoning. Such measures may includes eyetracking to find scanning patterns, as well as EEG, facial expressions and postural measures to infer and identify cognitive states associated with design processing (Bahr et al. 2007).

3. Data collection with experts and novices (undergraduate students) to capture rules of thumbs, individuality and variablity of reasoning processes. Critical data are how individuals, from expert to novice, solve a problem and their reasoning at the micro and marco levels; which alternatives designers conceive, as well as which solution is preferable and why. Moreover, a longitudinal study involving the same cohort of students acquiring expertise would provide considerably insight into the knowledge structures that support reasoning and how they change over time.

Having a foundation of reasoning knowledge data in place, the second step of the first challenge is modelling and storage of these data. Based on previous research (Hernandez et al. 2012, Bahr and Wood 2015, Bahr 2014) and the conceptual, organizational and associative affinity of relation systems with human long term memory, the development of a relational system is a possibility. The data will provide the foundation for modelling how knowledge of the experts is generally organized in line with cognitive psychological research. Data modelling is a rigorous process of iterative entity-relationship modelling, which results in a database schema. This schema may evolve over time but more importantly serves as the basic structure for organizing the semantic data acquired under the first challenge. An important aspect of the data model is the change of knowledge and reasoning over time. Hence, the data on individuality (across participants) and variability over time (from novice to expert) will be instrumental during the data modelling. As such a temporal data model will be necessary and similar to (Bahr and Wood 2015, Bahr 2014). Following the data modelling, the implementation into a relational database system seems to be a relatively trivial step and completes this challenge.

The result of challenge 1 is the construction of an artificial memory that includes the reasoning of a design domain. Reasoning is not just raw data but processed data. The set of research questions that relate to challenge 2 are based on the use and understanding of the data residing in the artificial memory using analytical and data mining techniques.

To investigate challenge 2, including schema extraction using data reduction techniques, expertise from the field of machine learning within artificial intelligence is necessary. In a first attempt the extraction and identification of relevant data from the different sub set of the artificial memory can be achieved by employing modern and well-established data mining tools like DMOP (Hilario et al. 2011) or KNIME (Berthold et al. 2008) to exploit the search space. A subsequent step could allow the examination of the previously mentioned dynamics and temporal components by applying other sophisticated methods like independent component analysis (Hyvärinen, Karhunen and Oja 2001) and recent deep learning approaches with convolutional neural networks and larger layer structures (Zeiler and Fergus 2014) that can be trained and efficiently applied by using the latest GPU accelerated graphics cards (Chetlur et al. 2014). Complementary or in addition, higher-level features might be identified by integrating sparse coding algorithms (Lee et al. 2006) that can also reduce the size of the data considerably to keep any processing steps manageable. Since a lot of parameters are involved in all stages of data handling, at least semi-automated optimization und fusion procedures appear as a necessity to allow maximizing the output of the various algorithmic classes. While taking all these different methods into account, the envisaged software architecture of the system should utilize design patterns (Gamma et al. 1995) from the field of software engineering to take care of sustainable aspects that support a long-term and continuous software development. This encompasses exemplarily categories like openness concerning the integration of source codes and methods from various different programming languages, flexibility allowing a change in the ordering of applied methods or working steps therein, and extensibility by providing software extension points that provide opportunities for the integration of customized modules via plug-in mechanisms (Ritter 2014). Having met challenges 1 and 2, the final and third challenge is the interaction design of a cognitive CAD tools that optimize the reasoning and semantic support during the design process.

Challenge 3 is the interaction design of a smart reasoning interface or intelligent CAD (iCAD) system. This challenge presents at least two research opportunities: how to make the artificial memory accessible to the design engineer during creative problem solving, and how to acquire new memories of reasoning and solutions alternatives during the design process.

The first effort requires interaction design that supports the user in an unobtrusive and “as needed basis” as to not interfere with (Bahr and Allen 2013) but enhance the primary task, i. e., the design process. This effort includes the use of good practices and principles of interaction design, such as spiral developments and techniques for early involvement of end-user in the interface design process (Bahr et al. 2006). Because it is intrinsic to the interaction that it occurs over time, each interaction creates a history of designs, which can be considered memories of problem solving. How these new memories are acquired and integrated is related the second research opportunity under the 3^rd challenge, the acquisition of new memories of reasoning and solution alternatives.

Such data may be gathered by explicitly prompting engineers for reasoning input during or after design completion. Moreover, these data may be inferred in post processing of the design steps taken by the engineer. These analyses are closely related to the research questions investigated under challenge 2 and are thus incremental in nature. In summary, this research requires the implementation of experimental ontologies that allow the system to learn from users and are able to self-populate with user data (i. e., solutions approaches). This captures some of the dynamic aspects intrinsic to expertise and begins to capture how knowledge grows and reasoning may change over time. In summary, supporting design engineers with cognitive CAD tools requires the design and implementation of a system that is a reflection of human memory and reasoning and that enhancing the CAD with these cognitive tools is a formidable task. In conclusion, the state of CAD as cognitive support tool for the design engineers is in its infancy and substantial cognitive psychological research and software development are still to be done.