Affiliations: [a] Department of Computer Science, University of Maryland, College Park, MD, USA | [b] Carnegie Mellon University, NASA Research Park, Mountain View, CA, USA
Correspondence:
[*]
Corresponding author: Timur Chabuk, Department of Computer Science, University of Maryland, College Park, MD, USA. E-mail: [email protected].
Abstract: In recent years, evolutionary computation has been successfully used to solve problems involving engineering design and invention, sometimes producing results that are qualitatively different than previous traditionally-designed solutions. However, while evolutionary methods appear to be a promising tool for supporting design, their usefulness is substantially limited by their computational expense and inability to integrate expert knowledge with evolutionary search. Here we develop and evaluate methods for causally-guided evolutionary design based on expert-supplied cause-effect relations that guide how genetic operators are applied (in contrast to conventional genetic operations which are carried out blindly and randomly), using these methods for antenna array design. To our knowledge, this is the first study that biases genetic operations in response to the specific performance characteristics of the individuals to which they are applied, and the first to use explicit cause-effect relations to guide this process. Our experimental evaluation compares using evolutionary systems with and without causal guidance to design directional dipole antenna arrays that meet pre-specified performance criteria. We find that causally-guided systems produce optimal solutions with significantly greater frequency and significant computational savings, suggesting that this approach may substantially improve the use of evolutionary computation in engineering design.
