Design Tools That Learn: A Possible CAD Future

Abstract

This paper introduces the concept of tools that learn. These are design support tools that acquire problem-specific knowledge as they are used to solve a problem. This knowledge is then reused in the solution of similar problems. The effect of this is that the tool is more efficient. An example is presented to demonstrate the idea. INTRODUCTION Computer-aided design has moved from meaning computer-aided drafting though computer-aided analysis to design synthesis support aids. Computer-aided drafting provides the means to document a completed or near completed design. It offered a number of advantages over manual drafting. Computer-aided analysis provides a means to determine the behaviours of a completed or partially complete design – you need to have a design before you can analyse it. In structural engineering design the development of matrix methods followed closely by the development of the finite element method provided a formal basis to the development of computer-aided analysis tools of considerable power. Other tools based on artificial intelligence techniques became available, tools such as code conformance checkers. Once analysis could be automated it became possible to develop design synthesis support aids. These tended to be aimed at parametric or routine design. In structural engineering we saw member design become automated. Once the context had been established the designs of frames and trusses could be automated. More recently we have seen other synthesis tasks having direct computer support. In general, where the design decisions are based on agreed evaluations such as code conformance and direct cost then human designers have allowed those design subtasks to be automated using design synthesis tools. This has produced what we might call ‘passive’ computer-based design aids, passive in the sense that the designer does not interact with them. Such design aids produce design decisions which most human designers would also produce since the tasks they deal with are sufficiently circumscribed to minimise the range of possible solutions which could be of interest given the agreed evaluators. In one sense these aids are the computational counterpart of design tables which were used extensively for member design previously. What all of these aids: computer-aided drafting tools, computer-aided analysis tools and ‘passive’ computer-based design aids have in common is that each tool is unchanged by its use. Every time to commence a new set of drawings on a computer-aided drafting system it takes the same effort. Every time you want to run a finite element analysis the program takes the same set of potential paths. Every time you want to synthesise a reinforced concrete column for a particular set of boundary conditions the program takes the same effort. There are clear and obvious benefits in having the tools unchanged by their use as this makes them independent of their use and they can be used with any arbitrary problem. However, I claim that there are also significant disadvantages in having the tools unchanged by their use. Each design that a designer works on adds to the experience of the designer, in this sense the designers learns from each design. For example, the designer in analysing a structure using finite element techniques may find that a certain type of element in a particular set of configurations produces much better results than the standard element. A designer working on a hospital layout using a computer-based layout planning aid may find that certain layouts of spaces produce better results than others. When each of these designers next tackles a similar design task it would be useful if the same tools now had knowledge about the fact that a certain type of finite element in a particular set of configurations produces much better results than the standard element and that certain layouts of spaces produce better results than others. The effect of this would be tools which are increasingly useful to the designer. In order for this to occur to tools would have to learn which are the beneficial aspects of the designs being produced. There is some relationship to this idea with case acquisition in case-based design systems (Maher at al 1995) and in systems which utilise constantly reinforced neural networks (Gunaratnam and Gero 1993).