Most of machine learning in design has focussed on learning generalizations to predict some future behavior of the system under consideration. Such approaches have been applied primarily to the analysis and synthesis stages of designing. There has been little work done relating to the formulation stage. This paper applies a particular machine learning approach to the improvement of the formal description of the design formulation. It applies an evolutionary technique to the problem reformulation to improve the formulation. This results in both a near optimal problem formulation and an improvement in the solution synthesized from that formulation.

This research originates in the work started several years ago at Worcester Polytechnic Institute dedicated to the investigation, modelling and evaluation of multiagent based design. The main thrust behind our approach is the idea of finding elementary patterns of agent problem-solving and interaction in design tasks. To achieve this goal we introduced and defined the concept of Single Function Agents, (SiFAs) (Dunskus, 1995; SiFA, 1995). SiFAs are agents specialized to perform one single generic function during the design process. Some typical functions would be selection, evaluation, and critique. These types of agents can be instantiated for different, particular design domains.

According to Simon (1983), machine learning is a process that improves the performance of an intelligent system. The performance task of a design system is to arrive at a structural description of a device given its functional description. Machine learning can occur, to improve the performance of the design system, in a number of ways. The learning can make the process of arriving at a known structure to function mapping faster, either by improving the control strategy or by learning new world knowledge needed for the structure to function mapping. It can also improve the performance of the design system by allowing it to come up with new structure to function mappings. The latter kinds of learning leads to innovative designs. The work presented in this paper, EnviroAdapt, belongs to the latter category. EnviroAdapt has the performance task of designing devices along with their operating environments. The EnviroAdapt learns about the novel environments and designs devices for these environments. This learning process is enabled by adapting the previous designs of devices along with their operating environments. This adaptation process makes use of abstracting general mechanisms from the previous designs, then instantiating them in the context of new design problem. Both of these processes are driven by the characteristics of the new design problem that requires a device to operate within a novel operating environment.

SALT is the highly influential system implemented by Marcus (1988), which is able to acquire knowledge and solve problems using the propose-and-revise strategy. The primary task to which this has been applied is the design of elevators, or lifts (Marcus et al., 1988). Recently we have re-implemented and extended the system at Aberdeen, S-SALT, the new system, has the following enhancements:

Most computer-based design tools assume designers work with a well-defined problem. The traditional treatment of design as two discrete phases; problem formulation and solution synthesis, is challenged by recent research. Though the view on discrete phases may be applicable to simple and/or well-defined design tasks, current research (Jonas, 1993; Logan & Smithers, 1993; Gero, 1994; Smithers et al., 1994) has shown that design is an ill-structured problem and the discrete phases view is not a good description of the process during which design alternatives are generated. A potential role of machine learning techniques is to provide a computational model of the changing representation of the design problem in response to the search for design solutions.

In this abstract, we discuss the use of learning techniques to improve performance and solution quality in multiagent parametric design. We have implemented the L-TEAM testbed for empirical evaluation of two forms of learning (described in detail below):

Though simulation models are extensively used for detailed design analysis, they find limited role in preliminary design decisions. We have developed a machine learning based approach to enable detailed simulation models to be harvested for supporting early-stage design of engineering systems.

Since the inception of research on machine learning (ML), these techniques have been associated with the task of automated knowledge generation or knowledge reorganization. This association still prevails, as seen in this issue. When the use of ML programs began to attract researchers in engineering design, different existing tools were used to test their utility and gradually, variations of these tools and methods have sprung up. In many cases, the use of these tools was based on availability and not necessarily applicability. When we began working on ML in design, we attempted to follow a different path (Reich, 1991a; Reich & Fenves, 1992) that led to the design of Bridger (Reich & Fenves, 1995), a system for learning bridge synthesis knowledge. Subsequent experiences and further reflection led us to conclude that the process of using ML in design requires careful and systematic treatment for identifying appropriate ML programs for executing the learning tasks we wish to perform (Reich, 1991b, 1993a). Another observation was that the task of creating or reorganizing knowledge for real design tasks is outside the scope of present ML programs. Establishing the practical importance of ML techniques had to start by addressing engineering problems that could benefit from present ML programs.

Creative design has been characterized in computational terms as “that design activity which occurs when a new variable is introduced into the design” (Gero, 1994). This is opposed to “routine design,” where “knowledge about variables, objectives expressed in terms of those variables, constraints expressed in terms of those variables, and the processes needed to find values for those variables, are all known a priori.” A third alternative, “innovative design,” occurs when no new variables are introduced, but when one or more variables are used with values outside the usual scope. In computational terms, routine design can also be seen as “search,” (sometimes also called “exploitation”) in so far as a certain, predefined search space is searched for a design solution. For creative design, where the search proceeds outside the boundaries of a predefined search space, the term “exploration” can be used.

We are working on using machine learning to make the numerical optimization of complex engineering designs faster and more reliable. We envision a system that learns from previous design sessions knowledge that enables it to assist the engineer in setting up and carrying out a new design optimization. We have performed initial experiments for two aspects of setting up an optimization: selecting a prototype to serve as a starting point for the optimization and selecting a reformulation of the search space. Both choices can dramatically affect the speed and the reliability of design optimization.