A Computer Program Automatically Acquiring Some Skills for a Simple Design Problem

One of the principal dimensions of expertise of a skilled designer is what has been called procedural know-how; experiential knowledge that enables goal-oriented problem solving. In this paper, we present an experiment in machine skill acquisition in a very small and narrow, but yet a demonstrative well-defined design domain through a program that acquires procedural know-how automatically by learning from its experience of practice (computational run) which it applies to produce good design solutions for a slightly different problem. The primitive learning and problem solving methods are based on an architecture for general intelligence called SOAR. The learned knowledge is analysed in terms of its contents, bounds of potential success and failure and why and how can it can be designated as a skill. What is surprising is that, a subset of the acquired know-how contains a very common human strategy used to generate the best solutions to this design problem.