Graph design by graph grammar evolution

摘要: Determining the optimal topology of a graph is pertinent to many domains, as graphs can be used model variety systems. Evolutionary algorithms constitute popular optimization method, but scalability concern with larger designs. Generative representation schemes, often inspired by biological development, seek address this facilitating discovery and reuse design dependencies allowing for adaptable exploration strategies. We present novel developmental method optimizing that based on notion directly evolving hypergraph grammar from which population derived. A multi-objective system established evaluated problems three domains: symbolic regression, circuit design, neural control. The observed performance compares favorably existing methods, extensive subgraphs contributes efficient solutions. Constraints also placed type explored spaces, ranging tree pseudograph. show more compact solutions are attainable in less constrained although convergence typically improves