This paper describes our preliminary results in applying constraint satisfaction techniques in a system we call TRANS-FORM for designing automatic automobile power transmissions. The work is being conducted in collaboration with the Ford Motor Company Advanced Transmission Design Department in Livonia, Michigan. Our current focus is on the design of the mechanical subsystem, but we anticipate extending this later to the electrical and hydraulic subsystems also. For simplicity, in the initial work reported here we restrict ourselves to the relatively well-explored class of transmissions having four forward speeds and one reverse speed, built from two planetary gearsets, cross-connected by two permanent links. Moreover, we pursue design of such transmissions only at the ‘kinematic level’. These two restrictions correspond to limiting respectively the breadth (generality) and the depth (detail or granularity) of the search space employed. We find that, at least for the restricted version of the problem pursued here, transmission design is an application very naturally formulated as a constraint satisfaction problem. Our present problem requires only 10 variables, with an average of about seven values each, and 43 constraints—making it similar in difficulty to about the 10-queens problem. So far, two of the classic transmissions, known as Axod and HydraMatic, have been rediscovered (at the kinematic level) by our program. Preliminary results also indicate that the constraint satisfaction framework will continue to remain adequate and natural even when the search space is allowed to be much broader and deeper. We expect that searches of such expanded spaces will soon lead to the discovery of totally new transmissions.

This paper describes the development of models to simulate the process of Concept Design Evaluation. The models are an amalgam of a number of statistically based methods and approaches taken from the probability, reliability, and quality domains. They assume that designers use decomposition of design to undertake evaluation at design characteristic level with the total design evaluation being achieved, in some way, via recomposition. The models described in this paper attempt to describe how designers may perform recomposition and hence total design evaluation. It is argued that the ability to model this human activity is important for the future development of knowledge-based design tools.

This paper focuses on that form of learning that relates to exploration, rather than generalization. It uses the notion of exploration as the modification of state spaces within which search and decision making occur. It demonstrates that the genetic algorithm formalism provides a computational construct to carry out this learning. The process is exemplified using a shape grammar for a beam section. A new shape grammar is learned that produces a new state space for the problem. This new state space has improved characteristics.

A main theme of concurrent engineering is the effective communication between relevant disciplines. Any computer tools for concurrent engineering must provide sufficient constructs and strategies for this purpose. This paper describes the AGENTS system, a domain-independent general-purpose Object-Oriented Prolog language for cooperating expert systems in concurrent engineering design. Emphasis is placed on demonstrating the use of the AGENTS constructs for distributed knowledge representation and the cooperation strategies for communication, collaboration, conflict resolution, and control. A simple case study is presented to illustrate the balance between simplicity and flexibility.

A framework IPD (Iterative Parametric Design) is proposed to assist the iterative parametric mechanical design process. To effectively find a set of satisfiable values for the design parameters the key is to find good heuristics to adjust or tune the parametric values resulting from previous design iterations. We propose that heuristics can come from two aspects by both qualitative and quantitative reasoning. Qualitative reasoning, based on confluences, provides global control over the feasible directions of variable adjustments, while quantitative reasoning, based on the dependency network and perturbation analysis, can be used to propose actual quantity of local variable adjustments. We used the design of a helical compression spring as an example to illustrate the performance of IPD system. We show that IPD can often find a solution faster than those without guidance of qualitative and quantitative reasoning.

This study focuses on the application of neural network techniques to adaptive remeshing of an idealized squareshaped structure and individual triangle by using triangular elements. The backpropagation learning algorithm is implemented by a supervised training technique to deal with the problem of remeshing structural elements in a structural analysis. A recent study introduced a structural remeshing that incorporated a finite-element analysis with an adaptive mesh generation technique. The main objective of the study is to demonstrate how neural networks can be employed to remesh structural elements without using numerically intensive computations. One essential requirement of this approach is the selection of feasible and appropriate training and testing data. The exploration of neural network in remeshing structural elements is a fundamental technique that looks beyond the finite-element and adaptive mesh generation techniques. It also demonstrates the capability of neural network to represent n-dimensional space and track each individual characteristic in that separate space. In general, an overview is presented, and the potential of neural networks to use the backpropagation algorithm instead of the more conventional approach of numerical methods is demonstrated.

The linking of research in machine learning with research in knowledge-based design is such that each of the two areas benefit from the consideration of the other. The use of machine learning in design addresses the perceived need to support the capture and representation of design knowledge, because handcrafting a representation is a difficult and time-consuming task. In addition, design provides a task with which to investigate the usefulness of existing machine learning techniques, and, perhaps, to discover new ones.

This paper defines, for use in design, rules for propagating “distribution constraints” through relationships such as algebraic or vector equations. Distribution constraints are predicate logic statements about the values that physical system parameters may assume. The propagation rules take into account “variation source causality”: information about when and how the values are assigned during the design, manufacturing, and operation of the system.

In this paper we present an application of AI search techniques to a class of problems that arise in transportation systems analysis. Rather than adapting the time-space network formulation typically used in Operations Research, we propose a discrete dynamic network to represent a scheduled service network. In a discrete dynamic network, there are finite, discrete, predetermined possibilities for moving from one vertex to another. Visiting a vertex has a cost (possibly zero), which may depend both on how the vertex was reached and how it will be left.

We describe the DYNET search algorithm for finding optimal paths in discrete dynamic networks. DYNET has been implemented in a working system (TRAINS) which searches the entire Dutch railway services network. An optimal path in a discrete dynamic network makes us arrive at our destination as early as possible (given our planned earliest departure time), and given this earliest arrival time (eat), will allow us to leave as late as possible, thereby guaranteeing a shortest path relative to the eat. DYNET first conducts a forward search to find the earliest possible arrival time, then a backward search which uses results of the forward search, to find the latest departure to arrive at that eat. Various AI techniques (symmetries, abstraction spaces, distance estimates, etc.) improve the performance of DYNET.

There exists a large body of Artificial Intelligence (AI) research on generating plans, i.e. linear or non-linear sequences of actions, to transform an initial world state to some desired goal state. However, much of the planning research to date has been complicated, ill-understood, and unclear. Only a few of the developers of these planners have provided a thorough description of their research products, and those descriptions that exist are usually unrealistically favorable since the range of applications for which the systems are tested is limited to those for which they were developed. As a result, it is difficult to evaluate these planners and to choose the best planner for a different domain. To make a planner applicable to different planning problems, it should be domain independent. However, one needs to know the circumstances under which a general planner works so that one can determine its suitability for a specific domain.

This paper presents criteria for evaluating AI planners; these criteria fall into three categories: (1) performance issues, (2) representational issues, and (3) communication issues. This paper also assesses four non-linear AI planners (NOAH, NONLIN, SIPE and TWEAK) based on a study of the published literature and on communication (via electronic mail, meetings and correspondence) with their developers.

The expertise of designers consists, primarily, of information about the relationship between goals or performance criteria and the attributes of the desired artifact that will result in performances that will satisfy these criteria. The designer like experts in other fields is typically better at applying the knowledge that constitutes his expertise than he is at articulating this knowledge. Generation and simulation models are discussed as a means of generating a set of designs for which the set of attributes defining these designs and the performance of these designs in terms of the criteria considered are explicitly defined. Pareto optimization is discussed as a means of structuring these designs on the basis of their performance. The induction algorithm ID3 is used as a means of inferring general statements about the nature of solutions which exhibit Pareto optimal performance in terms of a set of performance criteria. The rules inferred in building design domain are compared with those extracted using a heuristic based learning system.

The ability to understand the implications of the geometry of solid objects is an important aspect of intelligent behavior. This paper presents work designed to enable reasoning with relatively loose conceptualizations of geometry. The method operates by comparing target geometry to known geometry, and involves the manipulation of models based upon prototypes. In particular, three techniques of simplification, approximation, and transformation are discussed. Finally, an application of the method to the domain of stress concentration prediction is presented.

This issue of AIEDAM is based on a workshop on Machine Learning in Design held at the 1994 Conference on Artificial Intelligence in Design, AID'94, (Gero & Sudweeks, 1994); the second of such workshops, with the first being held at AID'92 in 1992 (Gero, 1992). The first workshop also resulted in a special issue of AIEDAM (Maher et al., 1994).