Knowledge that Transforms

The Spatial Adaptive Filter (SAF), introduced in this paper, uses generalized damped negative feedback to estimate spatially-varying parameters for multivariate models. Previous adaptive filters have been designed to estimate time-varying parameters and process data recursively in time sequence. SAF processes all data simultaneously in an iterative algorithm. Monte Carlo studies show that SAF is successful in automatically identifying and estimating step-jump and continuous spatial variation in the parameters of causal variables. A case study on census-tract data from Columbus, Ohio, relating police-vehicle hours spent in responding to calls to socio-economic indicators, has systematic spatial variation in estimated parameters. Independent variables that are significant in inner-city areas of Columbus become progressively less significant in moving to outlying areas.

This paper describes a study of Navy recruiter productivity at the individual recruiter level. Timeseries of monthly contract production by each of 345 recruiters who served for six months or more during the period May 1977–December 1978 formed the basis of the study. A wide variation in job tenure was observed in this data with some recruiters beginning their normal three-year tour of duty, and others in their second or third years. Two key empirical findings are reported here: (1) in addition to the expected learning period at the beginning of a recruiter's tour of duty there was a severe and extended demotivational (or “delearning”) period at the end of the tour. Production during these periods was very low. (2) substantial heterogeneity existed in recruiter productivity after controlling for the impact of tenure. Recruiter performance is predictable—good recruiters tend to stay good while poor recruiters continue to perform poorly. A stochastic model for heterogeneous production is developed and an early rotation policy for poorly performing recruiters is proposed. Under this policy a recruiter is observed for N months after the learning period. If total production in N months is less than c, the recruiter is rotated out and is replaced by a new recruiter. It is shown that the optimal selection of N and c can substantially improve recruiter force productivity. Productivity improves even when these parameters deviate somewhat from optimality. U.S. Navy Recruiting command initiatives that resulted from this and related studies are described. They included a change in the measure used for recruiter goal setting and an incentive system which led to rewards for superior recruiters and reassigning poorly performing recruiters to nonrecruiting duty. Finally the paper discusses the implications of the results obtained for industrial salesforces.

We consider a problem of sequencing capacity expansion projects with a continuous demand function specified over a given time horizon. Each type of expansion project has a specified integer capacity and an associated cost which is nonincreasing with respect to the time at which the project is brought on stream. The problem is to determine the sequence of expansions to provide sufficient capacity to meet demand at minimum cost. A formulation is presented and its relaxation leads to a shortest route problem. The sequencing problem is solved using a branch and bound procedure with Lagrangean relaxation providing bounds. A particularly effective heuristic is also developed. Computational results are given.

Component lot-splitting considerations, in which the lot-size of a component item may cover only a fraction of its parent item's lot-size, have been ignored in the literature when determining lot-sizes of items in multi-stage manufacturing systems. In this paper, a multi-stage lot-sizing problem is formulated under a specified component lot-splitting policy for the case of noninstantaneous production of items and constant demand for the end item. Optimal and heuristic solution procedures for the formulated problem are provided, including experimental results of comparison between these procedures. It is shown that considerable cost savings can result if the component lot-splitting approach is employed under favorable conditions in multi-stage manufacturing environments. In addition, reduced inventory levels are achieved which translate into lower working capital requirements and a less cluttered shop floor. The heuristic procedure is recommended as an acceptable alternative to the optimal procedure if the number of items in the multi-stage system is large or if inventory carrying and setup/order costs cannot be accurately estimated. Further, component lot-splitting considerations may be ignored if production rates of facilities in the system are in balance. Finally, a methodology for application of the component lot-splitting policy where the end item demand is time-varying is discussed.

In a recent paper we showed that unaided industry allocation to basic (inappropriable) research is suboptimal and that in stimulating this allocation, provision of government seed money is generally counterproductive, while the provision of matching subsidies is not cost-efficient. Here we consider a special case of the model developed in the earlier paper (i.e., we now consider a homogeneous industry) and investigate the effects of several relevant factors upon an industry’s allocation of resources to basic research. An extensive numerical example is presented that helps to verify and to interpret the model in realistic terms. Our findings question the validity of a number of popular beliefs about the need for government support of basic research in various types of industries. For example, contrary to popular belief, the greater the risk aversion displayed by member firms in an industry, the lesser may be the need for government support of its basic research. Also, the larger the number of firms in an industry the greater may be the need for government support of that industry's basic research.

Morris (Morris, P. A. 1983. An axiomatic approach to expert resolution. Management Sci. 29 24–32.) develops an axiom system for combining expert judgments by means of a processing rule. It is shown that the axioms are self-contradictory.

In this paper we present a heuristic method, based on Lagrangian relaxation, for multilevel lot-sizing when there is a single bottleneck facility. A series of Lagrangian relaxations (one for each item in the product structure) is imbedded in a branch and bound procedure. The objective is to find a production schedule that fits within available capacity at minimum cost. The method has two solution phases, dual and primal. In the dual phase of the procedure, implied costs of setups and production are determined based on a tentative schedule. The primal phase is repeated with these new prices and we iterate to reach a good solution. The solution procedure is first tested on two special cases: uncapacitated multilevel lot-sizing and the capacitated, single-level multi-item lot sizing problem. The results show that the solution procedure can provide better solutions than some heuristics designed especially for those problems. Test results on the bottleneck problem indicate that good feasible solutions are found for problems too difficult to solve with exact methods.

This paper describes a simulation procedure for estimating the distribution function of the shortest path length in a network with random arc lengths. The method extends the concept of conditional Monte Carlo utilizing special properties of the Uniformly Directed Cutsets and the unique arcs. The objective here is to reduce the sampling effort and utilize known probability information to derive multivariate integrals of lower dimension. The experimental results show that the proposed method is substantially cost effective and performs better than traditional Monte Carlo and conditional methods.

A theory based upon random search within a fixed population of technological possibilities is used to explain the manufacturing progress function. The theory is consistent with the power function relation between unit costs and cumulative output that has frequently been observed. It is also consistent with initial rates of improvement smaller than those predicted later by the power function relation, the eventual cessation of cost reduction, and an irregularity of improvements. Existing theories in the literature either fail to agree with the main empirical phenomena or else assume precisely what they attempt to explain.

The Herstein–Milnor independence axiom is a necessary condition for recursion analysis of decisions in extensive (=tree) form.