Julian J. Faraway

Time series forecasting with neural networks: a comparative study using the air line data

Summary. This case-study fits a variety of neural network (NN) models to the well-known airline data and compares the resulting forecasts with those obtained from the Box‐Jenkins and Holt‐ Winters methods. Many potential problems in fitting NN models were revealed such as the possibility that the fitting routine may not converge or may converge to a local minimum. Moreover it was found that an NN model which fits well may give poor out-of-sample forecasts. Thus we think it is unwise to apply NN models blindly in ‘black box’ mode as has sometimes been suggested. Rather, the wise analyst needs to use traditional modelling skills to select a good NN model, e.g. to select appropriate lagged variables as the ‘inputs’. The Bayesian information criterion is preferred to Akaike’s information criterion for comparing different models. Methods of examining the response surface implied by an NN model are examined and compared with the results of alternative nonparametric procedures using generalized additive models and projection pursuit regression. The latter imposes less structure on the model and is arguably easier to understand.

Regression analysis for a functional response

Functional responses are encountered when units are observed over time. Although the whole function itself is not observed, a sufficiently large number of evaluations, as is common with modern recording equipment, are assumed to be available. Functional regression analysis relates the smooth functional response, y(t), to known covariates, x, by a linear combination of parameter functions, β(t), which are to be estimated. The model takes the standard form, y(t) = x T β(t) + ∊(t). This approach provides an alternative to standard longitudinal data methods used in the biological sciences, where less and noisier data necessitate parametric modeling. The methodology is illustrated by an application in ergonomics.

Bootstrap Choice of Bandwidth for Density Estimation

Abstract A bootstrap-based choice of bandwidth for kernel density estimation is introduced. The method works by estimating the integrated mean squared error (IMSE) for any given bandwidth and then minimizing over all bandwidths. A straightforward application of the bootstrap method to estimate the IMSE fails because it does not capture the bias component. A smoothed bootstrap method based on an initial density estimate is described that solves this problem. It is possible to construct pointwise and simultaneous confidence intervals for the density. The simulation study compares cross-validation and the bootstrap method over a wide range of densities—a long-tailed, a short-tailed, an asymmetric, and a bimodal, among others. The bootstrap method uniformly outperforms cross-validation. The accuracy of the constructed confidence bands improves as the sample size increases.

Three Dimensional Analysis of Facial Movement in Normal Adults: Influence of Sex and Facial Shape

The aim of this study was to quantify facial movements in a sample of normal adults and to investigate the influence of sex and facial shape on these movements. The study sample consisted of 50 healthy adult subjects, 25 males and 25 females (age: mean = 27.3 years; range = 23-39 years). A video-based tracking system was used to track small-diameter retroreflective markers positioned at specific facial sites. Subjects were instructed to make 7 maximum facial animations from rest, and the facial movements for each animation were characterized as the vectors of maximum displacement. Hotellings T2 was used to test for significant sex differences in facial movements. In order to determine the effects of facial shape on facial movements, an index of facial shape was first calculated for each subject, and then a mixed-model ANOVA was used with facial shape (index), sex, and the interaction between facial shape and sex as fixed effects and subject as a random effect. The results demonstrated specific movement patterns for each animation. In general, males had larger movements than females and facial shape had a small but significant effect on facial movements. By comparing patient movements with the data from this large normative sample, the utility of this method to assess region-specific movement deficits was demonstrated.

Stature, Age, and Gender Effects on Reach Motion Postures

The rapid adoption of software to simulate human reach motions in the design of vehicle interiors and manufacturing and office workstations has required a sophisticated understanding of human motions. This paper describes how more than 3000 right-arm reaching motions of a diverse group of participants were captured and statistically modeled. The results demonstrate that stature and age have a larger effect than does gender on reach motion postures for motions chosen by the participants while reaching to targets placed throughout a typical automobile interior. We propose that these methods, models, and results can assist the further development of human motion simulation software for ergonomic purposes, such as for the design or evaluation of vehicle interiors or industrial workplaces, to ensure that various population groups are physically accommodated.

Rectifying postures reconstructed from joint angles to meet constraints

Postures are often described and modeled using angles between body segments rather than joint coordinates. Models can be used to predict these angles as a function of anthropometry and postural requirements. Postural representation, however, requires the joint coordinates. The use of conventional forward kinematics to derive joint coordinates from predicted angles may violate task constraints, such as the placement of a hand on a target or a foot on a pedal. Errors arise because the anthropometry or other motion characteristics of a subject, for which the prediction is to be made, may differ from the data from which the prediction model was derived. We describe how to rectify model-predicted postures to exactly satisfy such task constraints. We require that the model used for predicting the angles also produce estimates of the variation in these predictions. We show how to alter the initial angle predictions, with the amount of perturbation at each angle dependent on the accuracy of its estimation, so as to exactly satisfy the joint coordinate constraints. Finally, we show in an empirical example that this correction usually produces better overall predictions of posture than those obtained initially.

On the Cost of Data Analysis

Abstract A regression analysis usually consists of several stages, such as variable selection, transformation and residual diagnosis. Inference is often made from the selected model without regard to the model selection methods that preceeded it. This can result in overoptimistic and biased inferences. We first characterize data-analytic actions as functions acting on regression models. We investigate the extent of the problem and test bootstrap, jackknife, and sample-splitting methods for ameliorating it. We also demonstrate an interactive LISP-STAT system for assessing the cost of the data analysis while it is taking place.

The HUMOSIM Ergonomics Framework: A New Approach to Digital Human Simulation for Ergonomic Analysis

The potential of digital human modeling to improve the design of products and workspaces has been limited by the time-consuming manual manipulation of figures that is required to perform simulations. Moreover, the inaccuracies in posture and motion that result from manual procedures compromise the fidelity of the resulting analyses. This paper presents a new approach to the control of human figure models and the analysis of simulated tasks. The new methods are embodied in an algorithmic framework developed in the Human Motion Simulation (HUMOSIM) laboratory at the University of Michigan. The framework consists of an interconnected, hierarchical set of posture and motion modules that control aspects of human behavior, such as gaze or upper-extremity motion. Analysis modules, addressing issues such as shoulder stress and balance, are integrated into the framework. The framework encompasses many individual innovations in motion simulation algorithms, but the primary innovation is in the development of a comprehensive system for motion simulation and ergonomic analysis that is specifically designed to be independent of any particular human modeling system. The modules are developed as lightweight algorithms based on closed-form equations and simple numerical methods that can be communicated in written form and implemented in any computer language. The modules are independent of any particular figure model structure, requiring only basic forward-kinematics control and public-domain numerical algorithms. Key aspects of the module algorithms are “behavior-based,” meaning that the large amount of redundancy in the human kinematic linkage is resolved using empirical models based on laboratory data. The implementation of the HUMOSIM framework in human figure models will allow much faster and more accurate simulation of human interactions with products and workspaces using high-level, task-based control. INTRODUCTION Digital human figure models (DHM) are now widely used for ergonomic analysis of products and workplaces. In many organizations, DHM software is a tool of first resort for answering questions relating to physical interaction between people and objects. Yet any objective appraisal of the technology would conclude that the current reality of DHM software capability is far from the promise of a “digital human” that can interact realistically with products and environments. This paper is focused on efforts to improve the ability of DHM software to simulate physical posture and motion. Nearly every other aspect of DHM functionality also warrants improvement, including body shape representation, strength simulation, and cognitive function, but posture and motion are critical to the primary applications of DHM to the assessment of physical tasks. Posture simulation is as old as computerized manikins, because the manikin must be postured before an analysis can be conducted. Important early work was performed by Ryan for the U.S. Navy (Ryan 1970). Porter et al. (1993) summarized applications of digital human models in vehicle ergonomics during the early years of personal computers, at which time few of the current commercial DHM software tools were in use. Chaffin (2001) presented case studies of the expanding use of DHM for both product and workplace design and assessment. As evidence of the importance of posture and motion simulation, dozens of papers in the SAE literature and in other forums have presented a wide variety of methods for human simulating postures and motions, including multiple-regression (Snyder et al. 1972); analytic and numerical inverse kinematics (Jung et al. 1995; Tolani et al. 2000); optimization-based inverse kinematics (Wang and Verriest 1998); differential inverse kinematics (Zhang and Chaffin, 2000); functional regression on stretch-pivot parameters (Faraway 2000); scaling, warping, and blending of motion-capture data (Park et al. 2002; Faraway 2003; Monnier et al. 2003; Park et al. 2004; Dufour and Wang 2005); and many 2006-01-2365 The HUMOSIM Ergonomics Framework: A New Approach to Digital Human Simulation for Ergonomic Analysis Matthew P. Reed, Julian Faraway, Don B. Chaffin and Bernard J. Martin University of Michigan

Three-dimensional nasolabial displacement during movement in repaired cleft lip and palate patients

The objective of this study was two‐fold: (1) to explore the suitability of a novel modified Procrustes fit method to adjust data for head motion during instructed facial movements, and (2) to compare the adjusted data among repaired unilateral (n = 4) and bilateral (n = 5) cleft lip and palate patients and noncleft control subjects (n = 50). Using a video‐based tracking system, three‐dimensional displacement of 14 well‐defined nasolabial landmarks was measured during four set facial animations without controlling for head motion. The modified Procrustes fit method eliminated the contributions of head motion by matching the most stable landmarks of each video‐recorded frame of the face during function to frames at rest. Its effectiveness was found to approximate that of a previous method (i.e., use of a maxillary occlusal splint to which stable dentition‐based markers were attached). Data from both the unilateral and bilateral cleft lip and palate patients fell outside the normal range of maximum displacements and of asymmetry, and individual patients demonstrated greater right‐versus‐left asymmetry in maximum displacement than did individual noncleft subjects. It is concluded that the modified Procrustes fit method is fast, is easy to apply, and allows subjects to move the head naturally without the inconvenience of a splint while facial movement data are being collected. Results obtained using this method support the view that facial movements in cleft patients may be severely hampered and that assessment of facial animation should be strongly considered when contemplating surgical lip revisions. (Plast. Reconstr. Surg. 105: 1273, 2000.)

Sensitivity of a Method for the Analysis of Facial Mobility. I. Vector of Displacement

OBJECTIVE (1) To determine which facial landmarks show the greatest movement during specific facial animations and (2) to determine the sensitivity of our instrument in using these landmarks to detect putatively abnormal facial movements. DESIGN Movements of an array of skin-based landmarks on five healthy human subjects (2 men and 3 women; mean age, 27.6 years; range, 26 to 29 years) were observed during the execution of specific facial animations. To investigate the instrument sensitivity, we analyzed facial movements during maximal smile animations in six patients with different types of functional problems. In parallel, a panel was asked to view video recordings of the patients and to rate the degree of motor impairment. Comparisons were made between the panel scores and those of the measurement instrument. RESULTS Specific regions of the face display movement that is representative of specific animations. During the smile animation, landmarks on the mid- and lower facial regions demonstrated the greatest movement. A similar pattern of movement was seen during the cheek puff animation, except that the infraorbital and chin regions demonstrated minimal movement. For the grimace and eye closure animations, the upper, mid-facial, and upper-lip regions exhibited the greatest movement. During eye opening, the upper and mid-facial regions, excluding the upper lip and cheek, moved the most, and during lip purse, markers on the mid- and lower face demonstrated the most movement. We used the smile-sensitive landmarks to evaluate individuals with functional impairment and found good agreement between instrument rankings based on the data from these landmarks and the panel rankings. CONCLUSION The present method of three-dimensional tracking has the potential to detect and characterize a range of clinically significant functional deficits.