Knowledge that Transforms

We define a job-shop scheduling problem with three dynamic decisions: assigning due-dates to exogenously arriving jobs, releasing jobs from a backlog to the shop floor, and sequencing jobs at each of two workstations in the shop. The job-shop is modeled as a multiclass queueing network and the objective is to minimize both the work-in-process (WIP) inventory on the shop floor and the due-date lead time (due-date minus arrival date) of jobs, subject to an upper bound constraint on the proportion of tardy jobs. A general two-step approach to this problem is proposed: (1) release and sequence jobs in order to minimize the WIP inventory subject to completing jobs at a specified rate, and (2) given the policies in (1), set due-dates that will attempt to minimize the due-date lead time, subject to the job tardiness constraint. A simulation study shows that this approach easily outperforms other combinations of traditional due-date setting, job release, and priority sequencing policies for two cases (moderately loaded and heavily loaded) of a particular shop. As a result of the study, three scheduling principles are proposed that can significantly improve the performance of a two-station job-shop; in particular, better due-date performance can be achieved by ignoring due-dates on the shop floor. Although we have only considered a two-station shop, the approach and scheduling principles presented here might also be useful for larger shops.

Knowledge acquisition is a bottleneck for expert system design. One way to overcome this bottleneck is to induce expert system rules from sample data. This paper presents a new induction approach called CRIS. The key notion employed in CRIS is that nominal and nonnominal attributes have different characteristics and hence should be analyzed differently. In the beginning of the paper, the benefits of this approach are described. Next, the basic elements of the CRIS approach are discussed and illustrated. This is followed by a series of empirical comparisons of the predictive validity of CRIS versus two entropy-based induction methods (ACLS and PLS1), statistical discriminant analysis, and the backpropagation method in neural networks. These comparisons all indicate that CRIS has higher predictive validity. The implications of the findings for expert systems design are discussed in the conclusion of the paper.

Although economic activity between the U.S. and Japan has skyrocketed in the last decade, there are few large sample, cross-industry studies analyzing multiple forms of investment by the Japanese in the U.S. This study analyzes the key characteristics of each stage of industry evolution and the costs and benefits of each form of resource investment to predict the patterns of technological linkages, joint ventures, and direct investment of Japanese companies in the U.S. The results find support for a model predicting a predominance of technological linkages in emerging industries, joint ventures in growing industries, and direct investment in maturing industries. Technological linkages are most attractive in emerging industries as firms struggle to acquire technology, information and expertise and share cost and risk, yet retain flexibility. Joint ventures proliferate in growing industries because they offer a means of acquiring and expanding customer bases, yet reducing vulnerability. In maturing industries, where firms' key competencies are more developed, direct investment allows the company to generate demand in new markets without the disadvantage of joint governance.

This paper describes some rules for combining component models into complete linear programs. The objective is to lay the foundations for systems that give users flexibility in designing new models and reusing old ones, while, at the same time, providing better documentation and better diagnostics than is provided by current systems. The results presented here rely on two different sets of properties of LP models: first, the syntactic relationships among indices that define the rows and columns of the LP, and second, the meanings attached to these indices. These two kinds of information allow us to build a complete algebraic statement of a model from a collection of components provided by the model builder.

In this paper a branch-and-bound procedure is described for scheduling the activities of a project of the PERT/CPM variety subject to precedence and resource constraints where the objective is to minimize project duration. The procedure is based on a depth-first solution strategy in which nodes in the solution tree represent resource and precedence feasible partial schedules. Branches emanating from a parent node correspond to exhaustive and minimal combinations of activities, the delay of which resolves resource conflicts at each parent node. Precedence and resource-based bounds described in the paper are combined with new dominance pruning rules to rapidly fathom major portions of the solution tree. The procedure is programmed in the C language for use on both a mainframe and a personal computer. The procedure has been validated using a standard set of test problems with between 7 and 50 activities requiring up to three resource types each. Computational experience on a personal computer indicates that the procedure is 11.6 times faster than the most rapid solution procedure reported in the literature while requiring less computer storage. Moreover, problems requiring large amounts of computer time using existing approaches for solving this problem type are rapidly solved with our procedure using the dominance rules described, resulting in a significant reduction in the variability in solution times as well.

This paper examines the feasibility of rule-based forecasting, a procedure that applies forecasting expertise and domain knowledge to produce forecasts according to features of the data. We developed a rule base to make annual extrapolation forecasts for economic and demographic time series. The development of the rule base drew upon protocol analyses of five experts on forecasting methods. This rule base, consisting of 99 rules, combined forecasts from four extrapolation methods (the random walk, regression, Brown's linear exponential smoothing, and Holt's exponential smoothing) according to rules using 18 features of time series. For one-year ahead ex ante forecasts of 90 annual series, the median absolute percentage error (MdAPE) for rule-based forecasting was 13% less than that from equally-weighted combined forecasts. For six-year ahead ex ante forecasts, rule-based forecasting had a MdAPE that was 42% less. The improvement in accuracy of the rule-based forecasts over equally-weighted combined forecasts was statistically significant. Rule-based forecasting was more accurate than equal-weights combining in situations involving significant trends, low uncertainty, stability, and good domain expertise.

We study production planning problems where multiple item categories are produced simultaneously. The items have random yields and are used to satisfy the demands of many products. These products have specification requirements that overlap. An item originally targeted to satisfy the demand of one product may be used to satisfy the demand of other products when it conforms to their specifications. Customers' demand must be satisfied from inventory 100α% of the time. We formulate the problem with service constraints and provide near-optimal solution to the problem with fixed planning horizon. We also propose simple heuristics for the problem solved with a rolling horizon. Some of the heuristics performed very well over a wide range of parameters.