The formal practice of forecasting and planning F&P has risen to prominence within a few decades and now receives considerable attention from both academics and practitioners. This paper explicitly recognizes the nature of F&P as future-oriented decision making activities and, as such, their dependence upon judgmental inputs. A review of the extensive psychological literature on human judgmental abilities is provided from this perspective. It is argued that many of the numerous information processing limitations and biases revealed in this literature apply to tasks performed in F&P. In particular, the "illusion of control," accumulation of redundant information, failure to seek possible disconfirming evidence, and overconfidence in judgment are liable to induce serious errors in F&P. In addition, insufficient attention has been given to the implications of numerous studies that show that the predictive judgment of humans is frequently less accurate than that of simple quantitative models. Applied studies of F&P are also reviewed and shown to mirror many of the findings from psychology. The paper subsequently draws implications from these reviews and suggests reconceptualizing F&P through use of decision-theoretic concepts. At the organizational level this involves recognizing that F&P may perform many, often conflicting, manifest and latent functions which should be identified and evaluated through a multi-attribute utility framework. Operationally, greater use should be made of sensitivity analysis and the concept of the value of information.

Forecasting and Planning: An Evaluation

A 5-year-profiling in production mode at the University of Stuttgart has shown that more than 40% of the execution time of Message Passing Interface (MPI) routines is spent in the collective communication routines MPI_Allreduce and MPI_Reduce. Although MPI implementations are now available for about 10 years and all vendors are committed to this Message Passing Interface standard, the vendors’ and publicly available reduction algorithms could be accelerated with new algorithms by a factor between 3 (IBM, sum) and 100 (Cray T3E, maxloc) for long vectors. This paper presents five algorithms optimized for different choices of vector size and number of processes. The focus is on bandwidth dominated protocols for power-of-two and non-power-of-two number of processes, optimizing the load balance in communication and computation.

/pdf/optimization-of-collective-reduction-operations-3ods0mnch9.pdf

Optimization of Collective Reduction Operations

Many multiple criteria decision making methods have been proposed and applied to water planning. Their purpose is to provide information on tradeoffs among objectives and to help users articulate value judgments in a systematic, coherent, and documentable manner. The wide variety of available techniques confuses potential users, causing inappropriate matching of methods with problems. Experiments in which water planners apply more than one multicriteria procedure to realistic problems can help dispel this confusion by testing method appropriateness, ease of use, and validity. We summarize one such experiment where U.S. Army Corps of Engineers personnel used several methods to screen urban water supply plans. The methods evaluated include goal programming, ELECTRE I, additive value functions, multiplicative utility functions, and three techniques for choosing weights (direct rating, indifference tradeoff, and the analytical hierarchy process). Among the conclusions we reach are the following. First, experienced planners generally prefer simpler, more transparent methods. Additive value functions are favored. Yet none of the methods are endorsed by a majority of the participants; many preferred to use no formal method at all. Second, there is strong evidence that rating, the most commonly applied weight selection method, is likely to lead to weights that fail to represent the trade-offs that users are willing to make among criteria. Finally, we show that decisions can be as or more sensitive to the method used as to which person applies it. Therefore, if who chooses is important, then so too is how a choice is made.

Does choice of multicriteria method matter? An experiment in water resources planning

Those who conduct integrated assessments (IAs) are aware of the need to explicitly consider multiple criteria and uncertainties when evaluating policies for preventing global warming. MCDM methods are potentially useful for understanding tradeoffs and evaluating risks associated with climate policy alternatives. A difficulty facing potential MCDM users is the wide range of different techniques that have been proposed, each with distinct advantages. Methods differ in terms of validity, ease of use, and appropriateness to the problem. Alternative methods also can yield strikingly different rankings of alternatives. A workshop was held in which climate change experts and policy makers evaluated the usefulness of MCDM for IA. Participants applied several methods in the context of a hypothetical greenhouse gas policy decision. Methods compared include value and utility functions, goal programming, ELECTRE, fuzzy sets, stochastic dominance, min max regret, and several weight selection methods. Ranges, rather than point estimates, were provided for some questions to incorporate imprecision regarding weights. Additionally, several visualization methods for both deterministic and uncertain cases were used and evaluated. Analysis of method results and participant feedback through questionnaires and discussion provide the basis for conclusions regarding the use of MCDM methods for climate change policy and IA analyses. Hypotheses are examined concerning predictive and convergent validity of methods, existence of splitting bias among experts, perceived ability of methods to aid decision-making, and whether expressing imprecision can change ranking results. Because participants gained from viewing a problem from several perspectives and results from different methods often significantly differed, it appears worthwhile to apply several MCDM methods to increase user confidence and insight. The participants themselves recommended such multimethod approaches for policymaking. Yet they preferred the freedom of unaided decision-making most of all, challenging the MCDM community to create transparent methods that permit maximum user control. Copyright © 2002 John Wiley & Sons, Ltd.

An evaluation of multi-criteria methods in integrated assessment of climate policy

Many multiobjective analysis (MOA) methods differ greatly in what purpose they serve, their ease of use and validity, and the decisions they yield. This wide variety of available methods bewilders potential users, resulting in inappropriate matching of methods and problems and unnecessary user dissatisfaction. Experiments that apply different methods to the same problem can help dispell this confusion by clarifying how methods differ. Although such experiments have already been useful, better designed experiments that address neglected issues would teach us even more. Users of MOA methods are introduced to the knowledge that can be gained from MOA experiments. To illustrate the potentials and limitations of experiments, results from a comparison of MOA methods in power plant siting are summarized.

What Can We Learn from Experiments in Multiobjective Decision Analysis

Multiattribute utility models are used for evaluating alternatives when there are more than one criterion present. There is a trend toward the development of complicated versions of these models. These versions, although theoretically more accurate in the representation of decision makers' attitudes, require assessment procedures which are more difficult and time consuming to implement than simpler models. This paper reviews theoretical and empirical research involving the sensitivity analysis of multiattribute utility models in an attempt to answer the question of whether such additional complexities are worthwhile. Both deterministic and probabilistic models are considered and the studies are divided into four areas: (1) those involving sensitivity to the form of the multiattribute utility function; (2) those involving sensitivity to the parameters of the functions; (3) those involving sensitivity to the form and parameters of individual single attribute utility functions; and (4) those involving the relationship between deterministic and probabilistic models. A discussion of the results is given at the end.

Sensitivity analysis of the effect of variations in the form and parameters of a multiattribute utility model: A survey

This paper considers the architecture of clusters and related message-passing (MP) software algorithms and their effect on performance (speedup and efficiency) of cluster computing (CC). We present new architectures for multi-segment Ethernet clusters and new MP algorithms which fit these architectures. The multiple segments (e.g. commodity hubs) connect commodity processor nodes so as to allow MP to be highly parallelized by avoiding network contention and collisions in many applications where the all-gather and other collective operations are central. We analyze all-gather in some detail, and present new network topologies and new MP algorithms to minimize latency. The new topologies are based on a design, called two-by-four nets (2×4 nets ) , by Compbionics. An integrated MP software system, called Reduced Overhead Cluster Communication (ROCC), which embodies the MP algorithms is also described. In brief, 2×4 nets are networks of “supernodes”, called 2×4's, each having 4 processors on 2 segments and segments usually being Ethernet hubs. The supernodes are typically connected to form rings or tori of supernodes. We present actual test results and supporting analyses to demonstrate that 2×4 nets with the ROCC MP software are faster than many existing clusters and generally less costly.

Architectures and message-passing algorithms for cluster computing: design and performance

As a first step toward modeling the conditioning of the eye blink reflex in rabbit cerebellum, we have developed CAJAL, a general biological neural network simulator This paper describes the structure of CAJAL and also serves as a progress report of an initial version

CAJAL - 91: A Biological Neural Network Simulator

Continuing previous research, compartmental models of cerebellar networks were developed based on new data from extracellular recordings of cerebellar cortex and interpositus nucleus in rabbit during conditioning of the eyeblink response. Sites of synaptic plasticity were postulated, as suggested by these recordings and by recent cooling probe experiments, and these were incorporated into the models. The models were tested by simulation using the CAJAL simulator, which produces action potential spike trains that can be compared against the experimental time- trace data. Conditioning is simulated by varying parameters in the model synapses at the implicated sites. Results provide further evidence that these cerebellar circuits may be a main neural substrate of this conditioning paradigm.

Modeling and Simulation of Compartmental Cerebellar Networks for Conditioning of Rabbit Eyeblink Response

Electrophysiological recordings in visual cortex give evidence of oscillatory dynamics which undergoes a synchronization in response to certain correlated stimuli. We present two types of models of neural networks and show by analysis and computer simulation that these model networks have oscillatory dynamics which exhibits synchronization when subjected to inputs similar to the experimental stimuli. Effects of noisy inputs on synchronization are also simulated as well as chaotic oscillations.

P. K. Leung

Papers

Sensitivity analysis of the effect of variations in the form and parameters of a multiattribute utility model: A survey

Architectures and message-passing algorithms for cluster computing: design and performance

CAJAL - 91: A Biological Neural Network Simulator

Modeling and Simulation of Compartmental Cerebellar Networks for Conditioning of Rabbit Eyeblink Response

Analysis and Simulation of Synchronization in Oscillatory Neural Networks