Machine Learning is the study of methods for programming computers to learn. Computers are applied to a wide range of tasks, and for most of these it is relatively easy for programmers to design and implement the necessary software. However, there are many tasks for which this is difficult or impossible. These can be divided into four general categories. First, there are problems for which there exist no human experts. For example, in modern automated manufacturing facilities, there is a need to predict machine failures before they occur by analyzing sensor readings. Because the machines are new, there are no human experts who can be interviewed by a programmer to provide the knowledge necessary to build a computer system. A machine learning system can study recorded data and subsequent machine failures and learn prediction rules. Second, there are problems where human experts exist, but where they are unable to explain their expertise. This is the case in many perceptual tasks, such as speech recognition, hand-writing recognition, and natural language understanding. Virtually all humans exhibit expert-level abilities on these tasks, but none of them can describe the detailed steps that they follow as they perform them. Fortunately, humans can provide machines with examples of the inputs and correct outputs for these tasks, so machine learning algorithms can learn to map the inputs to the outputs. Third, there are problems where phenomena are changing rapidly. In finance, for example, people would like to predict the future behavior of the stock market, of consumer purchases, or of exchange rates. These behaviors change frequently, so that even if a programmer could construct a good predictive computer program, it would need to be rewritten frequently. A learning program can relieve the programmer of this burden by constantly modifying and tuning a set of learned prediction rules. Fourth, there are applications that need to be customized for each computer user separately. Consider, for example, a program to filter unwanted electronic mail messages. Different users will need different filters. It is unreasonable to expect each user to program his or her own rules, and it is infeasible to provide every user with a software engineer to keep the rules up-to-date. A machine learning system can learn which mail messages the user rejects and maintain the filtering rules automatically. Machine learning addresses many of the same research questions as the fields of statistics, data mining, and psychology, but with differences of emphasis. Statistics focuses on understanding the phenomena that have generated the data, often with the goal of testing different hypotheses about those phenomena. Data mining seeks to find patterns in the data that are understandable by people. Psychological studies of human learning aspire to understand the mechanisms underlying the various learning behaviors exhibited by people (concept learning, skill acquisition, strategy change, etc.).

Machine learning

According to W. Edwards Deming, American companies require nothing less than a transformation of management style and of governmental relations with industry. In Out of the Crisis, originally published in 1982, Deming offers a theory of management based on his famous 14 Points for Management. Management's failure to plan for the future, he claims, brings about loss of market, which brings about loss of jobs. Management must be judged not only by the quarterly dividend, but by innovative plans to stay in business, protect investment, ensure future dividends, and provide more jobs through improved product and service. In simple, direct language, he explains the principles of management transformation and how to apply them.

Out of the Crisis

As reflected in the amount of controversy, few areas in psychology have undergone such dramatic conceptual changes in the past decade as the emerging science of heuristics. Heuristics are efficient cognitive processes, conscious or unconscious, that ignore part of the information. Because using heuristics saves effort, the classical view has been that heuristic decisions imply greater errors than do “rational” decisions as defined by logic or statistical models. However, for many decisions, the assumptions of rational models are not met, and it is an empirical rather than an a priori issue how well cognitive heuristics function in an uncertain world. To answer both the descriptive question (“Which heuristics do people use in which situations?”) and the prescriptive question (“When should people rely on a given heuristic rather than a complex strategy to make better judgments?”), formal models are indispensable. We review research that tests formal models of heuristic inference, including in business organizations, health care, and legal institutions. This research indicates that (a) individuals and organizations often rely on simple heuristics in an adaptive way, and (b) ignoring part of the information can lead to more accurate judgments than weighting and adding all information, for instance for low predictability and small samples. The big future challenge is to develop a systematic theory of the building blocks of heuristics as well as the core capacities and environmental structures these exploit.

/pdf/heuristic-decision-making-4yjobmwu4i.pdf

Heuristic Decision Making

This paper is a review of research in product development, which we define as the transformation of a market opportunity into a product available for sale. Our review is broad, encompassing work in the academic fields of marketing, operations management, and engineering design. The value of this breadth is in conveying the shape of the entire research landscape. We focus on product development projects within a single firm. We also devote our attention to the development of physical goods, although much of the work we describe applies to products of all kinds. We look inside the "black box" of product development at the fundamentaldecisions that are made by intention or default. In doing so, we adopt the perspective of product development as a deliberate business process involving hundreds of decisions, many of which can be usefully supported by knowledge and tools. We contrast this approach to prior reviews of the literature, which tend to examine the importance of environmental and contextual variables, such as market growth rate, the competitive environment, or the level of top-management support.

/pdf/product-development-decisions-a-review-of-the-literature-1jhatri7bn.pdf

Product Development Decisions: A Review of the Literature

Past empirical work has shown that learning multiple related tasks from data simultaneously can be advantageous in terms of predictive performance relative to learning these tasks independently. In this paper we present an approach to multi--task learning based on the minimization of regularization functionals similar to existing ones, such as the one for Support Vector Machines (SVMs), that have been successfully used in the past for single--task learning. Our approach allows to model the relation between tasks in terms of a novel kernel function that uses a task--coupling parameter. We implement an instance of the proposed approach similar to SVMs and test it empirically using simulated as well as real data. The experimental results show that the proposed method performs better than existing multi--task learning methods and largely outperforms single--task learning using SVMs.

/pdf/regularized-multi-task-learning-13f3zxoz9e.pdf

Regularized multi--task learning

https://www.mit.edu/~hauser/Papers/VirtualCustomer.pdf

The virtual customer

http://web.mit.edu/edahan/www/WorkingPaperonInternetConceptTestingbyDahanandSrinivasan.PDF

The Predictive Power of Internet-Based Product Concept Testing Using Visual Depiction and Animation

We model concept testing in new product development as a search for the most profitable solution to a design problem. When allocating resources, developers must balance the cost of testing multiple designs against the potential profits that may result. We propose extreme-value theory as a mathematical abstraction of the concept-testing process. We investigate the trade-off between the benefits and costs of parallel concept testing and derive closed-form solutions for the case of profits that follow extreme-value distributions. We analyze the roles of the scale and tail-shape parameters of the profit distribution as well as the cost of testing in determining the optimal number of tests and total budget for the concept phase of NPD. Using an example, we illustrate how to estimate and interpret the scale and tail-shape parameters. We find that the impact of declining concept-testing costs on expected profits, the number of concepts tested, and total spending depend on the scale/cost ratio and tail-shape parameter of the profit distribution.

An Extreme-Value Model of Concept Testing

We propose and test new adaptive question design and estimation algorithms for partial profile conjoint analysis. Polyhedral question design focuses questions to reduce a feasible set of parameters as rapidly as possible. Analytic center estimation uses a centrality criterion based on consistency with respondents' answers. Both algorithms run with no noticeable delay between questions.

We evaluate the proposed methods relative to established benchmarks for question design random selection, D-efficient designs, adaptive conjoint analysis and estimation hierarchical Bayes. Monte Carlo simulations vary respondent heterogeneity and response errors. For low numbers of questions, polyhedral question design does best or is tied for best for all tested domains. For high numbers of questions, efficient fixed designs do better in some domains. Analytic center estimation shows promise for high heterogeneity and for low response errors; hierarchical Bayes for low heterogeneity and high response errors. Other simulations evaluate hybrid methods, which include self-explicated data.

A field test 330 respondents compared methods on both internal validity holdout tasks and external validity actual choice of a laptop bag worth approximately $100. The field test is consistent with the simulation results and offers strong support for polyhedral question design. In addition, marketplace sales were consistent with conjoint-analysis predictions.

/pdf/fast-polyhedral-adaptive-conjoint-estimation-56hbqfn2fk.pdf

Fast Polyhedral Adaptive Conjoint Estimation

Greedoid languages provide a basis to infer best-fitting noncompensatory decision rules from full-rank conjoint data or partial-rank data such as consider-then-rank, consider-only, or choice data. Potential decision rules include elimination by aspects, acceptance by aspects, lexicographic by features, and a mixed-rule lexicographic by aspects (LBA) that nests the other rules. We provide a dynamic program that makes estimation practical for a moderately large numbers of aspects.

We test greedoid methods with applications to SmartPhones (339 respondents, both full-rank and consider-then-rank data) and computers (201 respondents from Lenk et al. 1996). We compare LBA to two compensatory benchmarks: hierarchical Bayes ranked logit (HBRL) and LINMAP. For each benchmark, we consider an unconstrained model and a model constrained so that aspects are truly compensatory. For both data sets, LBA predicts (new task) holdouts at least as well as compensatory methods for the majority of the respondents. LBA's relative predictive ability increases (ranks and choices) if the task is full rank rather than consider then rank. LBA's relative predictive ability does not change if (1) we allow respondents to presort profiles, or (2) we increase the number of profiles in a consider-then-rank task from 16 to 32. We examine trade-offs between effort and accuracy for the type of task and the number of profiles.

http://web.mit.edu/hauser/www/Papers/Yee_Dahan_Hauser_Orlin_Greedoid_MKS_07.pdf

Ely Dahan

Papers

The virtual customer

The Predictive Power of Internet-Based Product Concept Testing Using Visual Depiction and Animation

An Extreme-Value Model of Concept Testing

Fast Polyhedral Adaptive Conjoint Estimation

Greedoid-Based Noncompensatory Inference