Optimal design of systems that evolve over time using neural networks

Abstract

Computational design optimization is challenging when the number of variables becomes large. One method of addressing this problem is to use pattern recognition to decrease the solution space in which the optimizer searches. Human "common sense" is used by designers to narrow the scope of search to a confined area defined by patterns conforming to likely solution candidates. However, computer-based optimization generally does not apply similar heuristics. In this thesis, a system is presented that recognizes patterns and adjusts its search for optimal solutions based on performance associations with these patterns. A design problem was selected that requires the optimization algorithm to assess designs that evolve over time. A small sensor network design is evolved into a larger sensor network design. Optimal design solutions for the small network do not necessarily lead to optimal design solutions for the larger network. Systems that are well-positioned to evolve have characteristics that distinguish themselves from systems that are not well-positioned to evolve. In this study, a neural network was able to recognize a pattern whereby flexible sensor networks evolved more successfully than less flexible networks.

(cont.) The optimizing algorithm used this pattern to select candidate systems that showed promise for successful evolution. In this limited exploratory study, a genetic algorithm assisted by a neural network achieved better performance than an unassisted genetic algorithm did. In a Pareto front analysis, the assisted genetic algorithm yielded three times the number of optimal "non-dominated" solutions as the unassisted genetic algorithm did. It realized these results in one quarter the CPU time. This thesis uses a sensor network example to establish the merit of neural network use in multi-objective system design optimization and to lay a basis for future study.