Learning Design Rules With Explicit Termination Conditions to Enable Efficient Automated Design

Abstract

We present a two-step technique for learning reusable design procedures from observations of a designer in action. This technique is intended for the domain of parametric design problems in which the designer iteratively adjusts the parameters of a design so as to satisfy the design requirements. In the first step of the two-step learning process, decision tree learning is used to infer rules that predict which design parameter the designer is likely to change for any particular state of an evolving design. In the second step, decision tree learning is again used, but this time to learn explicit termination conditions for the rules learned in the first step. The termination conditions are used to predict how large of a parameter change should be made when a rule is applied. The learned rules and termination conditions can be used to automatically solve new design problems with a minimum of human intervention. Experiments with this technique suggest that it can reproduce the decision making process observed from the designer, and it is considerably more efficient than the previous technique, which was incapable of learning explicit rule termination conditions. In particular, the rule termination conditions allow the new program to automatically solve design problems with far fewer iterations than previously required.