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 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.

The use of machine learning techniques requires the formulation of a learning problem in a particular domain. The application of machine learning techniques in a design domain requires the consideration of the representation of the learned design knowledge, that is, a target representation, as well as the content and form of the training data, or design examples. This paper examines the use of a target representation of design concepts and the application, adaptation, or generation of machine learning techniques to generate design concepts from design examples. The examples are taken from the domain of bridge design. The primary machine learning paradigm considered is concept formation.

One method for making analogies is to access and instantiate abstract domain principles, and one method for acquiring knowledge of abstract principles is to discover them from experience. We view generalization over experiences in the absence of any prior knowledge of the target principle as the task of hypothesis formation, a subtask of discovery. Also, we view the use of the hypothesized principles for analogical design as the task of hypothesis testing, another subtask of discovery. In this paper, we focus on discovery of physical principles by generalization over design experiences in the domain of physical devices. Some important issues in generalization from experiences are what to generalize from an experience, how far to generalize, and what methods to use. We represent a reasoner's comprehension of specific designs in the form of structure-behavior-function (SBF) models. An SBF model provides a functional and causal explanation of the working of a device. We represent domain principles as device-independent behavior-function (BF) models. We show that (1) the function of a device determines what to generalize from its SBF model, (2) the SBF model itself suggests how far to generalize, and (3) the typology of functions indicates what method to use.

Design knowledge is continually refined and expanded through experience. This research is concerned with design knowledge expressed as constraints. A simple learning mechanism simulates an expert designer's ability to incrementally adjust her knowledge when presented with slightly new problems. In response to unsatisfied expectations during the design process the system will examine its general knowledge about the design artifact, discover some relevant constraining knowledge, and convert that knowledge into a design constraint for future use. This process, referred to as constraint inheritance, should automatically improve the problem-solving performance.

The goal of machine learning for artifact synthesis is the acquisition of the relationships among form, function, and behavior properties that can be used to determine more directly form attributes that satisfy design requirements. The proposed approach to synthesis knowledge acquisition and use (SKAU) described in this paper, called NETSYN, creates a function to estimate the probability of each possible value of each design property being used in a given design context. NETSYN uses a connectionist learning approach to acquire and represent this probability estimation function and exhibits good performance when tested on an artificial design problem. This paper presents the NETSYN approach for SKAU, a preliminary test of its capability, and a discussion of issues that need to be addressed in future work.

This paper describes an application of artificial intelligence to support software reuse. We begin by discussing the characteristics of software engineering that establish dynamic reorganization as a requirement for a repository of software artifacts. We then present an experimental system that uses incremental concept formation as the basis for dynamic reorganization, and the conceptual hierarchy that was generated by the system for a set of 67 artifacts. The hierarchy is compared to a hierarchy produced manually by independent investigators, and the automatic hierarchy is evaluated in terms of retrieval efficiency and retrieval reliability. The paper ends with a discussion of three projects that share similar objectives with our work.