This issue of AIEDAM is based on a workshop on Machine Learning in Design held at the 1996 Conference on Artificial Intelligence in Design, AID'96 (Gero & Sudweeks, 1996), the third of such workshops, with the previous two being held at AID'92 (Gero, 1992) and at AID'94 (Gero & Sudweeks, 1994). The first two workshops also resulted in special issues of AIEDAM (Maher et al., 1994; Duffy et al., 1996).

This paper addresses an important application of machine learning (ML) in design. One of the major bottlenecks in the process of engineering analysis by using the finite-element method—a design of the finite-element mesh—was a subject of improvement. Defining an appropriate geometric mesh model that ensures low approximation errors and avoids unnecessary computational overhead is a very difficult and time-consuming task based mainly on the user's experience. A knowledge base for finite-element mesh design has been constructed using the ML techniques. Ten mesh models have been used as a source of training examples. The mesh dataset was probably the first real-world relational dataset and became one of the most widely used training set for experimenting with inductive logic programming (ILP) systems. After several experiments with different ML systems in the last few years, the ILP system CLAUDIEN was chosen to construct the rules for determining the appropriate mesh resolution values. The ILP has been found to be an effective approach to the problem of mesh design. An evaluation of the resulting knowledge base shows that the mesh design patterns are captured well by the induced rules and represent a solid basis for practical application. The aim of this paper is not only to present the real-life ML application to design, but also to describe and discuss a relation of the work being done to the topic of this special issue: the proposed “dimensions” of ML in design.

Learning is an intrinsic product of case-based reasoning. Acquiring new cases is one possible way of learning in a case-based system. These cases comprise mainly success knowledge. The successful cases are essentially used to generate new design solutions. But a case-based system also can make use of failure knowledge. In this paper we present how a case-based system can acquire failure cases for verification of the solution created by success cases. We describe IM-RECIDE, a system that uses case-based reasoning for solving design problems that are imperfectly described and explained. The learning aspect is focused and some of the machine learning dimensions in design are criticized. Experimental results in the domain of room configuration also are presented.

Many of the design systems developed in recent years incorporate some machine learning. The number of such systems already available, and the multitude of design learning opportunities that are slowly being revealed, suggest that the time is ripe to attempt to put these developments into a systematic framework. Consequently, in this paper we present a set of dimensions for machine learning in design research. We hope that it can be used as a guide for comparing existing work, and that it may suggest new directions for future exploration in this area.

The research and development of a simulation-based decision support system (SB-DSS) capable of assisting early collaborative design processes is presented. The requirements for such a system are included. Existing collaborative DSSs are shown to lack the capability to manipulate complex simulation-based relationships. On the other hand, advances within the machine learning in design community are shown to have a potential for providing, but have not yet addressed, simulation-based support for collaborative design processes. The developed SB-DSS is described in terms of its four principal components. First, the behavior-evaluation (BE) model is used to both structure individual, domain-specific decision models and organize these models into a collaborative decision model. Second, a probabilistic framework for the BE model enables management of the uncertainty inherent in learning and using simulation-based knowledge. Significantly, this framework provides a constraint satisfaction environment in which simulation-based knowledge is used. Third, a statistical neural network approach is used to capture simulation-based knowledge and build the probabilistic behavior models based on this knowledge. Fourth, since a probability distribution theory does not exist for the nonlinear neural network approaches, Monte Carlo simulation is introduced as a method to sample the trained neural networks and approximate the likelihoods of design variable values. Consequently, constraint satisfaction problem-solving capability is obtained. In addition, a mapping of the SB-DSS architecture onto a collaborative design agent framework is provided. Experimental evaluation of a prototype SB-DSS system is summarized, and performance of the SB-DSS with respect to search and usability metrics is documented. Initial results in developing the simulation-based support for collaborative design are encouraging. Lastly, a categorization of the machine learning approach and a critique of the proposed categorization scheme is presented.

Abstraction and generalization of layout design cases generate new knowledge that is more widely applicable to use than specific design cases. The abstraction and generalization of design cases into hierarchical levels of abstractions provide the designer with the flexibility to apply any level of abstract and generalized knowledge for a new layout design problem. Existing case-based layout learning (CBLL) systems abstract and generalize cases into single levels of abstractions, but not into a hierarchy. In this paper, we propose a new approach, termed customized viewpoint—spatial (CV–S), which supports the generalization and abstraction of spatial layouts into hierarchies along with a supporting system, SPIDA (SPatial Intelligent Design Assistant).

The application of machine learning (ML) to solve practical problems is complex. Only recently, due to the increased promise of ML in solving real problems and the experienced difficulty of their use, has this issue started to attract attention. This difficulty arises from the complexity of learning problems and the large variety of available techniques. In order to understand this complexity and begin to overcome it, it is important to construct a characterization of learning situations. Building on previous work that dealt with the practical use of ML, a set of dimensions is developed, contrasted with another recent proposal, and illustrated with a project on the development of a decision-support system for marine propeller design. The general research opportunities that emerge from the development of the dimensions are discussed. Leading toward working systems, a simple model is presented for setting priorities in research and in selecting learning tasks within large projects. Central to the development of the concepts discussed in this paper is their use in future projects and the recording of their successes, limitations, and failures.

Gradient-based numerical optimization of complex engineering designs offers the promise of rapidly producing better designs. However, such methods generally assume that the objective function and constraint functions are continuous, smooth, and defined everywhere. Unfortunately, realistic simulators tend to violate these assumptions, making optimization unreliable. Several decisions that need to be made in setting up an optimization, such as the choice of a starting prototype and the choice of a formulation of the search space, can make a difference in the reliability of the optimization. Machine learning can improve gradient-based methods by making these choices based on the results of previous optimizations. This paper demonstrates this idea by using machine learning for four parts of the optimization setup problem: selecting a starting prototype from a database of prototypes, synthesizing a new starting prototype, predicting which design goals are achievable, and selecting a formulation of the search space. We use standard tree-induction algorithms (C4.5 and CART). We present results in two realistic engineering domains: racing yachts and supersonic aircraft. Our experimental results show that using inductive learning to make setup decisions improves both the speed and the reliability of design optimization.

This paper presents a formalism for considering the issues of learning in design. A foundation for machine learning in design (MLinD) is defined so as to provide answers to basic questions on learning in design, such as, “What types of knowledge can be learnt?”, “How does learning occur?”, and “When does learning occur?”. Five main elements of MLinD are presented as the input knowledge, knowledge transformers, output knowledge, goals/reasons for learning, and learning triggers. Using this foundation, published systems in MLinD were reviewed. The systematic review presents a basis for validating the presented foundation. The paper concludes that there is considerable work to be carried out in order to fully formalize the foundation of MLinD.