Research

Preference elicitation, preference structures and assessment

Learning algorithms and mathematical models of crowdsourcing

Emotional design quantification: Aesthetics and proportionality, perception of sustainability, behavior modification through design

Decomposition and coordination strategies for large-scale systems, multidisciplinary design optimization (MDO)

Design for market systems: Enterprise-wide business, marketing, engineering, public policy and economic considerations

Analytical target cascading and analytical target setting

Optimal design of product platforms, portfolios, and product lines

Optimal design of system topologies

Monotonicity analysis

Global, parametric, mixed-discrete, and Pareto optimization

Distributed, multilevel, multidisciplinary system optimization

Artificial intelligence, expert systems and nonlinear mathematical optimization

Optimal design under uncertainty

Co-design: Combined optimal design and optimal control

Analytical craftsmanship

Architectural design

Automotive systems, especially hybrid and electric powertrains

Electromagnetic systems, especially antennas

Manufacturing and design integration

Structural design

Sustainable products and systems

How can “crowds” contribute useful information for design decisions?

Crowdsourced evaluation is an open innovation method for aggregating concept design evaluations from a crowd of possibly unknown evaluators. The hope is that expertise within the crowd can provide a useful crowd consensus evaluation. Reported successes in industrial case studies and the research literature deal only with "simple” tasks where expertise is not required. How can these successes be realized in “complex” engineering design tasks, where only a minority in the crowd has sufficient expertise?

Our current research combines concepts from human preference modeling, machine learning, and psychology to create mathematical models for crowdsourced aggregation when only a minority subset of the crowd are experts. In particular, we are interested in methods for filtering experts or at least derive theoretical bounds characterizing the limitations of crowdsourced evaluation.

For more information contact: Alex Burnap aburnap@umich.edu

Keywords: crowdsourcing, systems engineering, machine learning, preference modeling.

Crowdsourcing is explored as a possible alternative to conventional optimization techniques for solving tough design problems. Fatigue in answering detailed surveys and finding the right "crowd" for problems requiring specialized knowledge are major difficulties in such crowdsourcing. An alternative is to use online games as a data collection mechanism. Game playing involves extensive trial and error guided by human intuition, and the element of fun gives the required incentive for the crowd to keep providing inputs.

To test this idea, we developed a web-based game, EcoRacer, for an electric vehicle powertrain design and control problem. In this game, while very few players were able to find solutions close to the (known) optimum (computed through dynamic programming), players developed good heuristics quickly that can be used to constrain the solution space. The heuristics extracted from players can be used to enhance existing optimization algorithms for design and control and get close to optimum quickly.

For more information contact: A. Emrah Bayrak bayrak@umich.edu

Keywords: gamification, game with a purpose, EV, electric vehicle powertrain design.

Product or service attributes are often determined from different perspectives such as engineering, industrial design, operations, and public policy. Our research integrates these different design perspectives into a single decision- making framework. In particular, we have been working on electric vehicle (EV) and autonomous vehicle design problems by integrating electric powertrain simulation, autonomous fleet assignment simulation, charging station location network, consumer demand, and public policy models, into a large decision making framework. We compared the public investment impacts on EV markets in China and US, and extended this framework to autonomous vehicle sharing services. In addition, we are exploring how to make design decisions today by taking future long-term design evolution into account. We adopt the real option theory models used in financial engineering to hedge design risks and create smooth design transitions.

For more information contact: Namwoo Kang nwkang@umich.edu

Keywords: market systems, systems thinking, electric vehicle, autonomous vehicle.

A consumer preference model is a core part in design for market systems, capturing preferences on perceptual attributes. We examine the trade-off between forms and functions in a vehicle and its seats. We elicited consumers’ preferences on product design using parameterized 3D shapes and tested the effects of specific design variables on the perception of forms.

In addition, we are tackling two research questions for perceptual attributes. First, how do people recognize whether a product design has been copied from that of other products? By understanding this decision-making process, we develop a design methodology for firms to use almost identical physical attributes developed by incumbent firms in the market while designing a perceptually different product. The second question is whether optical illusions can change consumers’ behavior. We propose a methodology to design safer and more sustainable products/services through tweaking customers’ perception.

For more information contact: Namwoo Kang nwkang@umich.edu

Keywords: perceptual attributes, consumer preference, choice models, human-computer interaction.

Many factors go into the development of aesthetic styling for the next generation of a current successful design. Take for example the Cadillac CTS, the mid-sized sedan of a flagship brand with a storied history over 100 years in the making, yet using a relatively new design language (i.e., “Art and Design”). The next model year of this car must balance design cues and features that capture essential Cadillac qualities, yet “reach” enough to be appealing to future customers looking for a fresh design.

Our work in this area builds off collaborations with General Motors, to try and build a system that is able to mathematically capture a value of design reach with a value of brand recognition for new designs, yet abstract away the math by staying within the visual space to speak in the universal language of visual styling. Such as system is rooted deeply in psychology due to the perceptual nature of a vehicles representation, and builds off related work in crowdsourced aggregation and hierarchical representations from our other research projects. The aim of this system is to support design decisions already made by executives, design managers, and strategic brand managers during periodic “gate reviews” during the design process.

For more information contact: Alex Burnap aburnap@umich.edu, Yanxin Pan yanxinp@umich.edu

Keywords: brand recognition, design reach, human perception, preference modeling.

Design representation modeling of product form (e.g., aesthetics, styling, CAD) has a rich history in the design research community. Modeling approaches can be parametric with known functional forms (i.e., Bezier curves, polygon mesh control handles), nonparametric with constitutive building blocks such as shape grammars, and hybrid approaches combining elements of both. Common to these approaches is a “forward direction” from the design variables to the design representation, in which a free-form space has implicitly constraints defining what possible designs exist. For the case of a 2D image, this may be conceptualized as a space of pixel color values (red, green, blue) across a matrix of length and width. While the number of free-form states is very high (3^ #_pixels), the actual possible 2D designs exist on a much lower dimensional manifold within this space, perhaps defining what it means to be a passenger vehicle.

In this project, we take a large dataset of existing vehicles and project design attributes such as brand and body type to model the manifold defining the design space of possible 2D passenger vehicle product forms. In this manner, we extract the "inverse" function of what it means to be a passenger vehicle by estimating the “backward direction” from design representation to design variables. Moreover, we are “learning” (in computer science jargon) this function in an unsupervised manner using deep hierarchical generative models (i.e., deconvolutional neural network, convolutional autoencoder, generative adversarial network), allowing us to model to distribution defining the manifold itself. This allows us to create and morph new designs on the fly, so that designers can ask: What is the boundary between the archetypical design forms of, say, a Cadillac and an Audi?

For more information contact: Alex Burnap aburnap@umich.edu, Yanxin Pan yanxinp@umich.edu

Keywords: design automation, generative modeling, brand recognition, design reach.

There is much hype around the term “big data.” Nevertheless, there are many opportunities for inferring data-driven design decisions via large data sets and updated algorithms able to handle these data through recent advances in parallel CPU and GPU-processing architectures. We are interested in using databases made up of variables representing millions of previous customers and the products they actually purchased, and creating preference prediction models that can be tested using held-out portions of the database.

Our models are focused on “content-based” preference prediction rather than “collaborative filtering.” In other words, we look at variables describing you a person and the product as an identity that interacts with you meaningfully (e.g., with your personal values), and then try to predict future product purchases. Contrast this with the case of upvotes and downvotes from your favorite online music provider, in which just your upvote pattern history is enough to infer your preferences apart from the underlying variable correlations that are unknown.

The family of models we employ are referred to as "feature learning," and have shown impressive results by breaking records in the image and speech recognition fields (e.g., Siri voice recognition). These advances have been enabled by learning correlations amongst large swaths of raw data to build higher-level data that better capture the underlying phenomena. At the same time, the amount of market data generated relating customers and their purchases is growing exponentially. Can we learn more discriminative features amongst this market data so we can better predict which customer will prefer which design concept?

For more information contact: Alex Burnap aburnap@umich.edu, Yanxin Pan yanxinp@umich.edu

Keywords: feature learning, deep learning, preference modeling.

The aesthetic appeal of designed artifacts has been long recognized as significantly affecting customer preferences. To quantitatively capture these descriptive and predictive factors of aesthetic appeal and visual attention, we adopt the framework of design as a communication between designers and customers. This framework suggests that design communication occurs from designer to customer, hereafter referred to as forward communication. We extend this framework to include communication from customer to designer, or a backward communication direction of customer response. Motivated by a collaboration with practicing automotive designers, our research goal is to capture this forward and backward communication by identifying regions of a design that draw visual attention. We introduce a data-driven method using deep learning techniques to simultaneously quantify both the forward and the backward communication. This method does not require humans to directly provide attention data, instead this method predicts the attention region in the given design from humans’ feedback on its attribute values.

For more information contact: Yanxin Pan yanxinp@umich.edu

Keywords: design visual attraction, deep learning.

A key difficulty of design preference modeling is to link contributing factors of preference from multimodal data, which means data in multiple modalities. To capture that association, it is important to transform the original data representation into a new representation so that the information in different data modalities can be exchanged.

Previous preference modeling methods can only handle information from one data modality or rely on human-defined features, which are impractical due to the high time and labor costs. Multimodal representation learning methods developed in the machine learning community that can automatically learn a compact set of latent variables across multiple data modalities to capture the association have been successfully used in several applications. However, these methods haven’t been applied and validated in design preference modeling tasks that require managing heterogeneity. The research goal here is to develop quantitative models that simultaneously quantify both forward and backward communication from multimodal data. These models can be used to provide not only accurate prediction for design preference but also insights for improving the product and interpreting preferences. We hope to use the backward communication model to convert the mathematical patterns discovered in the forward communication model into a language that designers can understand and process.

For more information contact: Yanxin Pan yanxinp@umich.edu

Keywords: design preference modeling, multimodal data, feature learning

System integration remains a challenge in the design of large complex engineered systems causing large time and cost overruns. We hypothesize that interface problems between subsystems during integration is a major contributor to this challenge, and that these problems arise from incongruent boundaries between subsystems. The notion of incongruence refers to a lack of precision in subsystem boundary/interface definitions, for example, due to mismatched variables or levels of abstraction, leading to gaps or overlaps during embodiment design. These interface problems may be purely technical in nature, but the interaction between subsystems during design can be viewed from additional perspectives that involve humans: the individual designers, the teams within which they operate, and the organization that is comprised of those designers. We seek to quantitatively identify the sources of incongruence in complex systems design interfaces from both technical and human perspectives, and develop propose strategies to mitigate such incongruences.

For more information contact: Arianne Collopy acollopy@umich.edu

Keywords: systems design, interfaces, systems integration.

Designing, building, and managing Large-Scale Complex Engineered Systems (LaCES) such as modern global automobiles, nuclear power stations, or aerospace vehicles requires contributions from thousands of experts across numerous technical disciplines, often working over many years and with considerable geographical dispersion. Further complicating the design task, comprehensive knowledge of these systems is not accessible to any single designer, discipline, manager, or subsystem expert. Rather, systems-level and subsystems-level information is distributed among experts within each domain of expertise. The systems engineering discipline aims to address this situation, but there is increasing recognition that systems engineering must be more deeply informed by knowledge residing in traditionally non-engineering fields, including the behavioral and social sciences.

Successful design of LaCES depends largely on the effectiveness of the interactions among different discipline experts. The “designer” of a LaCES is actually a diverse, dispersed team of researchers, and thus the act of successful “designing” must be a collaborative, interdisciplinary cognitive effort. Insights from engineering practice and cognitive science might help inform how both individual and group cognition influence these interdisciplinary interactions in LaCES design: the thinking skills and group practices each engineer “brings to the table” shape the collaborative work, and a better understanding of these processes is necessary in order to improve the design and management of complex systems.

For more information contact: Melissa Greene greenemt@umich.edu

Keywords: systems design, cognition, interdisciplinarity.

Modularity is a well-known concept in product design. What does modularity mean and might it be successful for the specific case of ground vehicle fleets for the US Army and other services?

Successful modularity implementations have generally been accomplished through swappable mission equipment rather than large-scale vehicle and core components transformation. Concurrently, the Army Science and Technology community has demonstrated the technical feasibility of large-scale, transformative ground vehicle modularity, but the business case for modularity remains incomplete. Army leadership needs to assess confidently and holistically the right balance between modular and mission-specific (conventional) vehicle platforms while accounting for multiple criteria tradeoffs, including total lifecycle cost, mission utility, personnel requirements, and fleet adaptability.

In this project we develop a modeling and simulation environment to evaluate the strategic feasibility of a modular vehicle fleet and compare it to a baseline fleet considering various aspects of a fleet operation including manufacturing, transportation, performance, maintenance and personnel requirements. The current research phase considers only operational aspects of modularity assuming that technical feasibility for effective modularity is possible.

For more information contact: A. Emrah Bayrak bayrak@umich.edu

Keywords: modularity, modular vehicle design, adaptive vehicle fleet, product platforms, commonality.

Hybrid Electric Vehicles (HEVs) in the commercial market have a variety of powertrain architectures (topologies). Do these HEVs have the best powertrain architecture for their target use? Does a small sedan architecture fit a large truck?

Our earlier research has shown that vehicles for different purposes and duty cycles require different optimal powertrain topologies. In this project we use a "clean-sheet design" approach to avoid biasing design decisions and develop a rigorous framework for optimizing topology and energy management of HEV powertrains.

System-level optimization of hybrid powertrains must include component sizing and control strategies along with architecture design. We have created a general representation of hybrid architectures based on bond graphs which allows a single representation of the entire feasible design space of all possible architectures. This feasible space is disjoint space, thus posing challenges for system optimization. We employ a decomposition-based optimization strategy partitioning the problem into two levels, vehicle performance and feasibility, and coordinating the solution using Analytical Target Cascading.

For more information contact: A. Emrah Bayrak bayrak@umich.edu

Keywords: Clean-sheet design, bond graphs, ECMS, genetic algorithms, Analytical Target Cascading, design and control optimization, HEV, hybrid electric powertrain design

Decomposition-based optimization strategies are used for solving large complex engineered system design problems. Such strategies have two steps, (i) partitioning the system into smaller subsystems and (ii) coordinating the solution of the subsystem problems so the overall system solution can be achieved. Analytical Target Cascading (ATC) is a coordination strategy with good theoretical properties and we have used it successfully for designing commercial vehicle systems and hybrid vehicle architectures. ATC works as follows: The upper level master system starts with given top-level targets (system mission specifications) and cascades targets for the lower- level subsystems to achieve by solving a top-level optimization problem; the lower-level subsystems produce “responses” that aim to match these targets by solving their own local optimization problems. ATC iterates until all targets and responses match as closely as the problem constraints allow.

ATC can handle hierarchical or non-hierarchical coordination and allows sequential or parallel execution of subproblem solutions. Numerical experiments have shown that ATC with a parallel solution strategy often fails to achieve convergence, that is, consistency between targets and responses. The algorithm may oscillate, frieze, or even diverge, requiring a large number of iterations and high computational cost. This research is investigating how to achieve robust convergence strategies for such problems and coordination strategies.

For more information contact: Namwoo Kang nwkang@umich.edu

Keywords: multidisciplinary design optimization, analytical target cascading, parallel solving