The GREASE project is an investigation of the application of artificial intelligence to cutting fluid selection and blending for metal machining operations. The problem is to first diagnose the machining operations to determine what fluid characteristics are required, then to select a cutting fluid which satisfies the required characteristics. The problem is exacerbated by the need to select a single fluid to be used by multiple types of operations on a variety of materials. Diagnosis is relatively simple, but treatment specification is difficult due to the variety of operations to be handled.

GREASE uses heuristic search in which the evaluation function is heuristically constructed. The construction of the evaluation function begins with the determination of the characteristics of an optimal fluid based on deep knowledge of the machining operations and materials. This is then altered heuristically according to problems diagnosed with the current fluid. Once the evaluation function is complete, it is used to select an existing fluid from the product line. GREASE has been tested extensively with results which equal that of the experts and has been field tested by the Chevron Corporation.

This paper proposes an approach for solving the velocity and acceleration of the limited-dof (dof n < 6) parallel kinematic machines with linear active legs by means of translational/rotational Jacobian and Hessian matrices. First, based on the established or derived constraint and displacement equations, the translational/rotational Jacobian and Hessian matrices are derived. Second, the formulae for solving inverse/forward velocities and accelerations are derived from translational and rotational Jacobian/Hessian matrices. Third, a 2SPR + UPU PKM and a 2SPS + RPRR PKM are illustrated for explaining how to use this method. This approach is simple because it needs neither to eliminate 6-n rows of an n × 6 Jacobian matrix nor to determine the screw or pose of the constrained wrench.

The architecture and implementation of a mechanical designer's assistant shell called DEJAVU is presented. The architecture is based on an integration of design and CAD with some of the more well known concepts in case-based reasoning (CBR). DEJAVU provides a flexible and cognitively intuitive approach for acquiring and utilizing design knowledge. It is a domain independent mechanical design shell that can incrementally acquire design knowledge in the domain of the user. DEJAVU provides a design environment that can learn from the designer(s) until it can begin to perform design tasks autonomously or semi-autonomously. The main components of DEJAVU are a knowledge base of design plans, an evaluation module in the form of a design plan system, and a blackboard-based adaptation module. The existance of these components are derived from the utilization of a CBR architecture. DEJAVU is the first step in developing a robust designer's assistant shell for mechanical design problems. One of the major contributions of DEJAVU is the development of a clean architecture for the utilization of case-based reasoning in a mechanical designer's assistant shell. In addition, the components of the architecture have been developed, tailored or modified from a general CBR context into a more synergistic relationship with mechanical design.

A spatial reasoning system must be able to represent and manipulate the location of objects in its world model. There are two general schemes for doing this—absolute symbolic schemes based on an independent coordinate system, and relative schemes which reference objects in relation to other objects of known location. This paper is concerned strictly with the latter.

Given a relative scheme, there are several possible strategies for segmenting space. Two such strategies are identified and discussed. The first, called the situation-specific strategy is the one currently being explored and employed in most spatial reasoning systems. The second, here referred to as the general-purpose or cognitive strategy is the one used by the human cognitive system. It is suggested that while both strategies have outstanding strengths and weakness, the latter holds greater potential for achieving maximal coverage with minimal resources. The paper then proceeds to describe the structure of cognitive models of locative space and to specify how such models can be built from 3-D geometric models.

This description is based on a cognitively motivated implementation called SEE-TELL. SEE-TELL takes as its input a 3-D geometric model and outputs a proposition of the form locative (referent, relatum). The function which maps the input to the output is a two part heuristic procedure. The first part determines the referent and relatum and the second part assigns the locative predicate. The system can assign the predicates on, right-of/left-of, and front-of/back-of. On and right-of/left-of are, respectively, illustrative of invariant and variant locatives. Front-of/back-of allows for a conflict between an ego/observer and a canonical object. These three situations are thought to cover the different classes of problems that can arise in assigning any locative.

The paper concludes by summarizing the findings, identifying the shortcomings and limitations, and making suggestions for future work.

Computer programs that combine traditional numeric methods with symbolic algebra and with specific knowledge of application-based techniques can provide new levels of computational support for engineering design. We illustrate this with a computer-based ‘control engineer’s assistant’. Although this program is focussed on control system design, it demonstrates techniques that should be widely applicable across many engineering disciplines. In particular, we show how, with symbolic computing, a computer-aided design system can usefully simulate engineering models early in the design process, before all (or any) system parameters have been specified numerically. Our system employs a flexible, extensible, object-oriented representation for control systems, which admits multiple mathematical models of designs and provides a framework for integrating tools that operate on diverse representations.

This paper defines, develops algorithms for, and illustrates the design use of a class of mathematical operations. These operations accept as inputs a system of linear constraint equations, Ax = b, an interval matrix of values for the coefficients A, and an interval vector of values for either x or b. They return a set of values for the other variable that is “sufficient” in this sense. Suppose that ◯ is an interval of input vectors, and  an interval matrix. Then, one Sufficient-Points operation returns a set of vectors ~ such that for each b in ~, the set of x values that can be produced by inserting all the values of  into Ax = b is a superset of the input vector x. These operations have been partly overlooked by the interval matrix mathematics community, but are mathematically interesting and useful in the design, for example, of circuits.

This paper describes OARPLAN, a prototype planning system that generates construction project plans from a description of the objects that comprise the completed facility. OARPLAN is based upon the notion that activities in a project plan can be viewed as intersections of their constituents: objects, actions and resources. Planning knowledge in OARPLAN is represented as constraints based on activity constituents and their interrelationships; the planner functions as a constraint satisfaction engine that attempts to satisfy these constraints. The goal of the OARPLAN project is to develop a planning shell for construction projects that (i) provides a natural and powerful constraint language for expressing knowledge about construction planning, and (ii) generates a facility construction plan by satisfying constraints expressed in this language.

To generate its construction plan, OARPLAN must be supplied with extensive knowledge about construction objects, actions and resources, and about spatial, topological, temporal and other relations that may exist between them. We suggest that much of the knowledge required to plan the construction of a given facility can be drawn directly from a three-dimensional CAD model of the facility, and from a variety of databases currently used in design and project management software. In the prototype OARPLAN system, facility data must be input directly as frames. However, we are collaborating with database researchers to develop intelligent interfaces to such sources of planning data, so that OARPLAN will eventually be able to send high level queries to an intelligent database access system without regard for the particular CAD system in which the project was designed.

We begin by explaining why classical AI planners and domain specific expert system approaches are both inadequate for the task of generating construction project plans. We describe the activity representation developed in OARPLAN and demonstrate its use in producing a plan of about 50 activities for a steel-frame building, based on spatial and topological constraints that express structural support, weather protection and safety concerns in construction planning. We conclude with a discussion of the research issues raised by our experiments with OARPLAN to date.

One of the products of engineering, besides constructed artifacts, is design documentation. To understand how design participants use documentation, designers and typical documentation users were interviewed and protocols were taken of them both creating and using design documentation. The protocols were taken from realistic projects of preliminary design for heating, ventilation, and air conditioning systems (HVAC). The studies of document creation and use revealed three important issues: (1) design participants not only look up design facts; they frequently access documents to obtain information about the rationale for design decisions; (2) the design rationale that they see often is missing from the documents; and (3) design requirements change frequently over a project life cycle so that design documents are often inconsistent and out of date. Recognizing these documentation issues in design practice, a new approach was developed in which documents are no longer static records, but rather interactive design models supporting a case. The feasibility of the approach was demonstrated by constructing a running system and testing it with designers on realistic problems. The costs and benefits of creating and using documentation of design rationale also were analyzed. In particular, the active documents approach was evaluated for a routine, preliminary design in domains where community practice is widely shared and largely standardized. The approach depends on the feasibility of creating a parametric design model for the design domain.

The following erratum is to correct an omission of a table during the publication phase of the article entitled: “User interface for specification language for case-based mechanical design” by Abhay Dandekar, Ibrahim Zeid, and Theodore Bardasz (AI EDAM, 11(1), p. 18).

The table referred to reads:

The Editor-In-Chief and Publisher regret the inadvertent mistake.

In this report the architecture of an intelligent design machine capable of performing routine design in different design domains is described. This machine is crafted to operate as a part of a larger system driven by a human designer. Notable features are its use of best-first search strategies for problem-solving control and its ability to adjust problem-solving control strategies, perform automated redesign following specification changes, and resolve constraint violations using domain knowledge. The claims made for this machine are analyzed and it is argued that these claims are founded on established principles of design for intelligent systems. An implementation of this architecture in a rule-based system named Proteus is discussed and its operation is examined using as an example the domain of register-transfer-level computer design.

This paper presents a review of CBD and its application to building design in particular. Case-based design is the application of case-based reasoning to the design process. Design maps well to case-based reasoning because designers use parts of previous design solutions in developing new design solutions. This paper identifies problems of case representation, retrieval, adaptation, presentation, and case-based maintenance along with creativity, legal, and ethical issues that need to be addressed by CBD systems. It provides a comprehensive review of CBD systems developed for building design and provides a detailed comparison of the CBD systems reviewed.

Pareto optimality is a domain-independent property that can be used to coordinate distributed engineering agents. Within a model of design called Redux, some aspects of dependency-directed backtracking can be interpreted as tracking Pareto optimality. These concepts are implemented in a framework, called Next-Link, that coordinates legacy engineering systems. This framework allows existing software tools to communicate with each other and a Redux agent over the Internet. The functionality is illustrated with examples from the domain of electrical cable harness design.

The function of a process planning system is to determine the methods by which a product is to be manufactured economically and competitively. In a modern manufacturing environment, a process planning system consists of highly trained people and complex software. The plans prepared by a process planning system are not always executed as planned. It is useful if the system can discover why plans fail, when they do fail. In order to learn why plans fail, the system must analyse a number of plans, both successful and unsuccessful, to find patterns in the failures of plans. This type of analysis is difficult for people, who are much better at analysing single events than multiple events.

The aim of the project described here is to design and implement a computer program which will help human planners in a process planning system to understand why plans fail. To achieve this aim, a program called IMAFO (Intelligent MAnufacturing FOreman) has been developed. IMAFO uses decision tree induction to analyse examples of both successful and unsuccessful plans.

The difficulties presented by this application are discussed and solutions are presented. Problems addressed include finding an appropriate set of attributes for describing the plans, using data efficiently, consolidating input from distinct sources, and presenting decision trees in an understandable form. Potential applications and directions for future research are considered.

A multifaceted approach to evaluating expert systems is overviewed. This approach has three facets: a technical facet, for “looking inside the black box”; an empirical facet, for assessing the system’s impact on performance; and a subjective facet, for obtaining users’ judgments about the system. Such an approach is required to test the system against the different types of criteria of interest to sponsors and users and is consistent with evolving lifecycle paradigms. Moreover, such an approach leads to the application of different evaluation methods to answer different types of evaluation questions. Different evaluation methods for each facet are overviewed.

This paper describes work in progress aimed at developing an interactive modeling tool that assists engineers with the task of physical modeling in finite element analysis. Physical modeling precedes the numerical simulation phase of finite element analysis and involves applying modeling idealizations to real world physical systems so that complex engineering problems are more amenable to numerical computation. In the paper, the nature of physical modeling is explored, a cognitive model of how engineers are thought to model complex problems is described and based on this model a knowledge-based modeling assistant is proposed. The AI approach taken is based on Chandrasekaran's propose-critique-modify design model adapted for the task of physical modeling. Within this framework, the AI paradigms of case-based reasoning, derivational analogy and model-based reasoning are exploited. By representing fundamental thermal modeling scenarios as cases, complex physical systems can be modeled in a piecewise fashion. Derivational analogy permits generative adaptation of retrieved cases by using model-based engineering traces thereby providing a basis for critiquing case solutions. An initial prototype is described which has been implemented for the domain of convection heat transfer analysis.

In a construction project, resource leveling techniques are necessary as a primary schedule-improvement tool to reduce overall project cost by decreasing day-to-day fluctuation in resource usage and resource idleness. There are, however, some limitations in traditional resource leveling techniques. Conventional heuristic approaches cannot guarantee a near-optimum solution for every construction project; a given heuristic may perform well on one project and poorly on another. The existing optimization approaches, such as linear programming and enumeration methods, are best applicable only to small size problems. Recently, there has been success in the use of Artificial Neural Networks (ANNs) for solving some optimization problems. The paper discusses how state-of-the-art ANNs can be a functional alternative to traditional resource leveling techniques. It then investigates the application of different ANN models (such as backpropagation networks, Hopfield networks, Boltzmann machines, and competition ANNs) to resource leveling problems. Because the development of ANNs involves not only science but also experience, the paper presents various intuitive yet effective ways of mapping resource leveling problems on different applicable ANN architectures. To demonstrate the application of ANNs to resource leveling, a simple ANN model is developed using a Hopfield network. The conclusion highlights the usefulness and the limitations of ANNs when applied to resource leveling problems.