Search

This report presents the functional specifications and top-level constructs of the core design of an image understanding (IU) application development environment. The core environment presented in this report provides tools to leverage the development of IU applications and to facilitate transfer of IU and reasoning technology from its origins in research laboratories into IU applications. The core environment provides both a development platform and reusable components including a library of image processing and IU routines and data structures, and an integrated set of high-level reasoning capabilities such as bayesian networks and logic engines. The environment design is an object oriented structure built on the C+ + programming language. The report also addresses system engineering issues in applying the environment to develop workstations specialized for terrain analysis and medical applications. Vision environment, Image Understanding (IU), Computer Vision (CV), object oriented programming.

This handbook is intended to provide decision makers and their staffs (current or potential) with an introduction to the basic concepts and operations of decision analysis. Decision analysis is a quantitative method which permits the systematic evaluation of the costs or benefits accruing to courses of action that might be taken in a decision problem. It entails identification of the alternative choices involved, the assignment of values (costs/benefits) for possible outcomes, and the expression of the probability of those outcomes being realized. With this information at hand, one can then systematically combine the values and probabilities to show the probable gain or loss that is associated with alternative choices.

Adversary surveillance has long been a concern of naval commanders at sea. Not only can it identify a naval task force as a possible military target, but it may also lead to an unwanted confrontation with possibly far greater military and political consequences. Ships at sea are potentially subject to many kinds of surveillance. Traditionally, the important components of a surveillance system have been aircraft, shipping, submarines, land-based radar, and communications sensors. In recent years, however, surveillance systems have added land-based acoustic sensors and earth-orbiting satellites with varied sensing capabilities.

This report describes Phase V efforts in the continuing investigation of the theoretical foundations of the decision-analytic model developed for the Office of Naval Research's Operational Decision Aiding Project. The principal issue under consideration was the degree to which the utilization of an incomplete, simple hierarchical model in the Tactical Decision Aid (TACAID) could lead to errors in inference as compared with a more complex Bayesian hierarachical model. Specifically, the more complex model would be designed to make allowances for data dependencies as well as nonstationarity in enemy intent. It was concluded that a simple model can lead to significant and potentially costly errors in inference, and that TACAID should be modified to handle these additional complexities. Concurrent with such modifications, training in the proper use of the more complex TACAID would be mandatory to ensure proper use of the additional capabilities. As part of the current research and developement effort, a situation definition capability, called TACDEF, was implemented for TACAID.

Recent technological advances have made possible sophisticated data- gathering capabilities for earth-orbiting spacecraft. Naval forces would appear to be particularly vulnerable to satellite surveillance since they operate on the surface of uniform, low-noise oceans where camouflaging is difficult. A methodology is presented for helping a naval commander avoid satellite surveillance. With known satellite orbital characteristics, the time and location of detection zones along a ship's planned route can be calculated. It is suggested that a commander can avoid many of these detection zones by varying his speed as a function of time. A graphical representation is presented and its limitations are discussed. A decision aid based on this methodology has been developed. It uses probabilistic information, such as weather, to find the expected number of detections for any speed profile. An optimal speed profile is found by means of dynamic programming.

Bayesian hierarchical inference is an optimum methodology for diagnosing the probable state of an unknown variable from data which are known, but which are known, but which are only indirectly related to the variable of interest. The present report describes a general procedure for Bayesian hierarchical inference implemented on an interactive graphics computer. It allows an intelligence analyst to build a hierarchical inference tree structure, assess the required probability distributions, and compute the likelihood distribution over the target variable. The procedure allows variables to be either, discrete, such as the type of research being pursued by a foreign government, or continuous, such as the personnel strength of a foreign tank regiment. (Author)

One strategy to improve the effectiveness of intelligence analysis is to provide decision makers with intelligence products which have been derived from a formal methodology such as decision analysis, and which include the logic of the methodology so that decision makers will better understand the reasoning of the intelligence analysis as well as their conclusions. In order to implement this strategy, however, methodological aids must be developed which are transparent to the analyst, and easy for him to use. The research described here represents an attempt to provide effective and easy to use decision analysis aids through the use of an interactive graphics computer. A series of prototype procedures have been developed which allow an intelligence analyst to make direct or indirect assessments of continuous probabilities, to calibrate his assessments with respect to discrete probabilities, to analyze the effects of future changes in force compositions on the strategic arms balance, and to discover the implications of intelligence information on policy decisions. Since the success fo these prototype procedures cannot be judged until intelligence analysts have had an opportunity to use them in their daily work, it is recommended that they be calculated in an operating environment. (Author)

This report describes a methodology for the application of a multi-attribute utility model to negotiations. Using a hypothetical base rights negotiation as an example, the paper describes how each party's preferences for each issue are assessed in terms of weighted utility curves, the set of Pareto optimum treaties is calculated, the value to each side of not having a treaty is assessed, and the treaty which is best with respect to four very general criteria is found. The question of disclosure is then explored. It is found that one side may increase the value of a treaty in some negotiations by falsifying its reported utility curves and that the presence of a mediator is, therefore, essential if the best possible treaty is to be found.

The report describes research in decision theory and computer aids for decision analysis. Summaries are presented on theoretic concepts including procedures for incorporating multiple dimensions of value in a single measure of utility, improving the structure of decision models, identifying cross cultural differences in the perception of uncertainty and optimizing the amount of information disclosure during bargaining. Several case study applications investigated use of decision analysis for assessment and forecasting, resource allocation, multi-dimensional evaluation, policy formulation and implementation, and negotiation modeling., See also report dated 24 May 1974, AD779861

Early efforts to produce decision aids, although they led to valuable research findings, did not succeed as field applications. Even when the targeted users agreed that the computer-aided decision process did improve decision making, these aids eventually fell into disuse when the personal support of the decision analysts was removed. User attitudes tended to be dominated by the short-range considerations of speed, convenience, and simplicity of operation, rather than by the long-range factors such as methodological thoroughness and consistency. The primary goal of this research project was to develop a prototype version of a user-friendly decision aid which could be used by persons with little or no prior training. This decision aid was not only designed to be easy to use, but also to save time and fit into the user's office or field site, while still providing the benefits of a technically sound decision-analytic approach. Since then, a prototype system has been developed, and a working model has been delivered to DARPA. This report is intended to accompany that working model. Its function is to document the purpose of the research effort, the technical approach, accomplishments of the project, and the prognosis for future development based on lessons learned during the effort. Because of the aid's self-explanatory nature and its unique hardware configuration, neither a users guide nor a system manual is needed.

Six psychological theories, from three areas of psychology, are analyzed with respect to the essential characteristics of the basic normative model for behavior in decision theory, the maximization of expected value. Despite differences in interpretation and level of analysis, five of the theories are shown to be formally equivalent to expected value models; that is they postulate that behavior is determined by two independent components corresponding to the probability that possible behaviors will have various outcomes, and the worth of these outcomes to the organism; and the performed behavior will be the one with the maximum net average worth. The sixth theory is interpretable as a weighted additive utility model. (Author)

The paper begins by pointing out the logical similarities between Expected Value (EV) theory and a number of theoretical formulations that have had successful careers in diverse areas of psychology. The notion is introduced that the failures of EV in contrast to the successes of the other theories may reside in the overprecision of the predictions and measurements in EV research as opposed to the rough predictions (e.g., more than, less than) and measurements (e.g., frequency response ratings on 7-point scales) in research deriving from the other theories. Then data are introduced from three experiments that imply that subjects are more tolerant of variance in their responses than EV experimenters are and that it, therefore, may be futile to try to expect highly precise predictions to be supported by the subjects' imprecise responses. (Author)

The paper describes the normative-descriptive approach to cognitive research. In a set of illustrative experiments based on models from physics, subjects were asked to judge the positions and weights necessary to maintain equilibrium in mechanical displays of varying complexity. Whenever possible subjects' judgments were accounted for through the appropriate normative models, otherwise descriptive models were generated from the normative models and the results of the other experiments. (Author)

An experiment in decision making examined the degree to which personal probabilities, subjective probabilities which need not have objective counterparts, conform to probability theory. Subjects gave subjective probability estimates for complex and elementary events. The degree of conformity to probability theory was evaluated by the degree to which probabilities for complex events could be predicted from the proper combination of the probabilities for elementary events. (Author)