Incorporating expert knowledge into evolutionary algorithms with operators and constraints to design satellite systems

Abstract

Knowledge-intensive evolutionary algorithms (EA) and knowledge-driven optimization algorithms that leverage problem- or domain-specific knowledge have been shown to discover high-quality solutions with fewer function evaluations than knowledge-independent EAs. Knowledge-intensive EAs apply the available knowledge through (1) knowledge-dependent operators that modify specific decision variables by biasing them towards or away from predetermined values or (2) knowledge-dependent constraints that penalize or eliminate solutions that are inconsistent with the provided knowledge. While both knowledge-dependent operators and knowledge-dependent constraints have shown promise in improving the search performance of an EA, there are no known experiments comparing their efficacy. To address the lack of such comparative experiments, this paper benchmarks one EA using knowledge-dependent operators and two EAs using knowledge-dependent constraints against an analogous knowledge-independent EA on a design problem for a climate-monitoring satellite system. Each EA is evaluated for its ability to attain high-quality solutions with the fewest possible number of function evaluations. In addition, we analyze the extent to which each method can focus on applying the knowledge that improves the search performance, handle conflicting information that suggests improving solutions with opposing modifications, and balance the exploitation of the available knowledge with the exploration of the tradespace to prevent premature convergence on local optima. The results show promise for EAs applying knowledge-dependent operators with an adaptive operator selection strategy and reveal some limitations of methods that adaptively apply knowledge-dependent constraints.