Evolving artificial neural networks with generative encodings inspired by developmental biology

摘要: In this dissertation I investigate the difference between generative encodings and direct for evolutionary algorithms. Generative are inspired by developmental biology were designed, in part, to increase regularity of synthetically evolved phenotypes. Regularity is an important design principle both natural organisms engineered designs. The majority focuses on how property enables a encoding outperform controls, whether bias towards also hurts performance some problems. report researchers can types regularities produced accommodate user preferences. Finally, study degree which produces another principle, modularity. Several previous studies have shown that highly regular However, prior dissertation, it was not known compare problems with different levels regularity. On three problems, show exploit intermediate amounts problem regularity, enabled increasingly controls as increased. This gap emerged because artificial neural networks (ANNs) behaviors. ANNs contained diverse array complicated, wiring patterns, whereas control irregular. I document hurt amount irregularity. propose new algorithm, called HybrID, wherein patterns modifies those provide fitness-enhancing irregularities. HybrID outperformed alone nearly all raises question may ultimately excel stand-alone algorithms, but being hybridized further process irregular refinement. The results described so far produce solutions. then that, at least case study, possible influence produced, allows domain knowledge preferences be injected into algorithm. investigated modular present first documented producing phenotype simple problem. encoding's inability create modularity harder where would been beneficial suggests more work needed likelihood response challenging, decomposable Overall, paints complete picture than studies. general conclusion drawn from properties seen complex, organisms, will likely part our long-term goal evolving phenotypes approach capability, intelligence, complexity their rivals.