Madan Mohan Dabbeeru

Discovering implicit constraints in design

Abstract Designers who are experts in a given design domain are well known to be able to Immediately focus on “good designs,” suggesting that they may have learned additional constraints while exploring the design space based on some functional aspects. These constraints, which are often implicit, result in a redefinition of the design space, and may be crucial for discovering chunks or interrelations among the design variables. Here we propose a machine-learning approach for discovering such constraints in supervised design tasks. We develop models for specifying design function in situations where the design has a given structure or embodiment, in terms of a set of performance metrics that evaluate a given design. The functionally feasible regions, which are those parts of the design space that demonstrate high levels of performance, can now be learned using any general purpose function approximator. We demonstrate this process using examples from the design of simple locking mechanisms, and as in human experience, we show that the quality of the constraints learned improves with greater exposure in the design space. Next, we consider changing the embodiment and suggest that similar embodiments may have similar abstractions. To explore convergence, we also investigate the variability in time and error rates where the experiential patterns are significantly different. In the process, we also consider the situation where certain functionally feasible regions may encode lower dimensional manifolds and how this may relate to cognitive chunking.

The Birth of Symbols in Design

In the widespread endeavour to standardize a vocabulary for design, the semantics for the terms, especially at the detailed levels, are often defined based on the exigencies of the implementation. In human usage, each symbol has a wide range of associations, and any attempt at definition will miss many of these, resulting in brittleness. Human flexibility in symbol usage is possible because our symbols are learned from a vast experience of the world. Here we propose the very first steps towards a process by which CAD systems may acquire symbols is by learning usage patterns or image schemas grounded on experience. Subsequently, more abstract symbols may be derived based on these grounded symbols, which thereby retain the flexibility inherent in a learning system. In many design tasks, the “good designs” lie along regions that can be mapped to lower dimensional surfaces or manifolds , owing to latent interdependencies between the variables. These low-dimensional structures (sometimes called chunks ) may constitute the intermediate step between the raw experience and the eventual symbol that arises after these patterns become stabilized through communication. In a multi-functional design scenario, we use a locally linear embedding (LLE) to discover these manifolds, which are compact descriptions for the space of “good designs”. We illustrate the approach with a simple 2-parameter latch-and-bolt design, and with a 8-parameter universal motor.Copyright

Learning Concepts and Language for a Baby Designer

We introduce the “baby designer enterprise” with the objective of learning grounded symbols and rules based on experience, in order to construct the knowledge underlying design systems. In this approach, conceptual categories emerge as abstractions on patterns arising from functional constraints. Eventually, through interaction with language users, these concepts get names, and become true symbols. We demonstrate this approach for symbols related to insertion tasks and tightness of fit. We show how a functional distinction - whether the fit is tight or loose - can be learned in terms of the diameters of the peg and the hole. Further, we observe that the same category distinction can be profiled differently - e.g. as a state (clearance), or as a process (the act of insertion). By having subjects describe their experience in unconstrained speech, and associating words with the known categories for tight and loose, the frequencies of words associated with these can be discriminated. The resulting linguistic labels learned show that for the state profile, the words “tight” and “loose” emerge, and for the action, we get “tight” and “easy”. Once an initial grounded symbol is available, it is argued that knowledge-based systems based on such symbols can be sanctioned by its semantics, as well as its syntax, leading to more flexible usage.

Grounded discovery of symbols as concept-language pairs

In human designer usage, symbols have a rich semantics, grounded on experience, which permits flexible usage - e.g. design ideation is improved by meanings triggered by contrastive words. In computational usage however, symbols are syntactic tokens whose semantics is mostly left to the implementation, resulting in brittle failures in many knowledge-based systems. Here we ask if one may define symbols in computational design as {label,meaning} pairs, as opposed to merely the label. We consider three questions that must be answered to bootstrap a symbol learning process: (a) which concepts are most relevant in a given domain, (b) how to define the semantics of such symbols, and (c) how to learn labels for these so as to form a grounded symbol. We propose that relevant symbols may be discovered by learning patterns of functional viability. The stable patterns are information-conserving codes, also called chunks in cognitive science, which relate to the process of acquiring expertise in humans. Regions of a design space that contain functionally superior designs can be mapped to a lower-dimensional manifold; the inter-relations of the design variables discovered thus constitute the chunks. Using these as the initial semantics for symbols, we show how the system can acquire labels for them by communicating with human designers. We demonstrate the first steps in this process in our baby designer approach, by learning two early grounded symbols, tight and loose.

FUNCTIONAL PART FAMILIES AND DESIGN CHANGE FOR MECHANICAL ASSEMBLIES

We consider two questions related to functional part families: a) how to characterize function in a computational framework, and b) how does the structure-to-function model generalize when the design changes, e.g. by changing the set of design variables? For the first, we observe that function is defined on the space of behaviours of the part, whereas structure is defined in the space of design parameters. For mechanical assemblies, as the design parameters change, their effect on the motion parameters can be complex, and cannot be automated in full generality. Thus, the mapping from structure-to-function involves considerable designer knowledge. For computational purposes, we quantify this function by defining part-family-specific Configuration Space (C-space) constructions, and also a metric that operates on these C-spaces to define each function. When the design is changed, either by changing the design space (structure), or by the user expectation (function), can existing design knowledge from the earlier part family migrate to the new product family? We make a start towards exploring how this knowledge can be modified when the part family is evolved, for example by introducing additional design variables, or by changing functional roles. Using examples from several lock designs, we present a small prototype for this process of modeling function and design change, implemented on a commercial CAD engine.Copyright

Multi-objective functional analysis for product portfolio optimization

Product portfolio optimization requires that one identify a few designs that provide for the widest variety of functions while minimizing product variety. This requires one to identify groupings of products that meet different functional tradeoffs. Here we propose to use multi-objective optimization for estimating the non-dominated sets of designs, and mapping these to the design space reveals that the good designs are often restricted to a few patches on a low-dimensional manifold, thus resulting in significant dimensionality reductions for the design decision space. We model function in a design family in terms of a phenomenological description, leading to a set of performative behaviours at the functional level, which are determined using set of performance metrics specific to a given embodiment. The non-dominated designs are clustered in the design space in an unsupervised manner to obtain candidate product groupings which the designer may inspect to arrive at portfolio decisions.We demonstrate this process on two different designs (faucets and springs), involving both continuous and discrete design variables. The effect of numerical stability in the process is investigated empirically, and the conditions under which the results would scale to large dimensional spaces are also explored.

Computational Models of Tacit Knowledge

When an expert designer uses a term such as “interference fit” or “H7-r6”, they effortlessly invoke a rich set of associations across a wide range of experience. While at one level, the meaning of a term such as H7 is formally specified, many of these associations are implicit and hard to characterize formally. The explicit concepts build on layers of implicit abstraction; e.g. the concept of fit would be difficult to achieve without the commonsense notion of “tight”, discriminated by human infants from five month onwards. We propose that such ubiquitous expertise may be acquired as functionally relevant low-dimensional chunks in an experiential space, which are then stabilized through language. The technical terms of design build on these everyday concepts by mechanisms such as extension or narrowing of their semantics. We suggest a two-stage computational analog of this process: (a) the baby designer stage learns elementary concepts as tacit patterns on an input space; and (b) the novice designer stage relates these early concepts to explicitly defined design terms to arrive at a grounded semantics for the new symbols. We illustrate the process through the development of concepts such as interference fit.

Product Portfolio Selection of Designs Through an Analysis of Lower-Dimensional Manifolds and Identification of Common Properties

Functional commonalities across product families have been considered by a large body of product family design community but this concept is not widely used in design. For a designer, a functional family refers to a set of designs evaluated based on the same set of qualities; the embodiments and the design spaces may differ, but the semantics of what is being measured (e.g., strength of a spring) remain the same. Based on this functional behaviour we introduce a product family hierarchy, where the designs can be classified into phenomenological design family, functional part family and embodiment part family. And then, we consider the set of possible performances of interest to the user at the embodiment level, and use multi-objective optimisation to identify the non-dominated solutions or the Pareto-front. The designs lying along this front are mapped to the design space, which is usually far higher in dimensionality, and then clustered in an unsupervised manner to obtain candidate product groupings which the designer may inspect to arrive at portfolio decisions. We highlight and discuss two recently suggested techniques for this purpose. First, with help of dimensionality reduction techniques, we show how these clusters in low-dimensional manifolds embedded in the high-dimensional design space. We demonstrate this process on three different designs (water faucets, compression springs and electric motors), involving both continuous and discrete design variables. Second, with the help of a data analysis of Pareto-optimal solutions, we decipher common design principles that constitute the product portfolio solutions. We demonstrate this so-called ‘innovization’ principles on a spring design problem. The use of multi-objective optimisation (evolutionary and otherwise) is the key feature of both approaches. The approaches are promising and further research should pave their ways to better design and manufacturing activities.

Using Symbol Emergence to Discover Multi-Lingual Translations in Design

Incorporating design knowledge into computational design requires “symbols” — but this term as used in knowledge-based models of design is a formal term, defined only in terms of other symbols. For most humans, symbols are [term : meaning] pairs that emerge while interacting with real designs. However, both the term and its interpretation vary considerably across design groups, particularly in today’s international cooperative design scenario. For translating symbols in design, one needs to incorporate the design context, which is since the actual design object and its characteristics form the most relevant part of the context. In this work, we consider an embodied symbols approach towards translation, where models corresponding to symbol semantics are discovered based on functional norms in a given design context. The functions are available as performance measures on a given task, and lead to low-dimensional characterizations (called image schema ) that reveal inter-relations in the input space that must hold for functional validity. Some of these image schemas eventually acquire language labels and become symbols. Since different designers differ in experience and in language their symbols differ somewhat. Here we consider how independent language agents may map these low-dimensional characterizations (called chunks) to units of languages based on human commentary produced in the same context. We demonstrate how this process may work for the simple domain of insertion tasks and fits, and learn both the image schemas and the language labels in two different languages, English and Telugu.Copyright

Product Platform Selection in Lower-Dimensional Manifold Spaces

Product portfolios need to present the widest coverage of user requirements with minimal product diversity. User requirements may vary along multiple performance measures, comprising the objective space, whereas the design variables constitute the design space, which is usually far higher in dimensionality. Here we consider the set of possible performances of interest to the user, and use multi-objective optimization to identify the non-domination or the pareto-front. The designs lying along this front are mapped to the design space; we show that these “good designs” are often restricted to a much lower-dimensional manifold, resulting in significant conceptual and computational efficiency. These non-dominated designs are then clustered in the design space in an unsupervised manner to obtain candidate product groupings which the designer may inspect to arrive at portfolio decisions. With help of dimensionality reduction techniques, we show how these clusters in low-dimensional manifolds embedded in the high-dimensional design space. We demonstrate this process on two different designs (springs and electric motors), involving both continuous and discrete design variables.Copyright