Application of evolutionary algorithms for the design of neural network architectures

U današnje vrijeme, neuronske mreže sve su veći predmet interesa. Oponašanjem funkcionalnosti i povezanosti neurona u mozgu, umjetne neuronske mreže mogu učiti te raditi predikcije. Razvijene su mnoge tehnike učenja, a u ovom je radu opisano učenje genetskim algoritmom, još poznato i kao neuroevolucija. Implementirane su dvije neuronske mreže. U prvoj je genetski algoritam korišten samo za optimizaciju težina, dok je u drugoj NEAT algoritam primijenjen za optimizaciju same arhitekture neuronske mreže. Te su neuronske mreže korištene za rješavanje problema regresije. Primjer rada neuronskih mreži prikazan je na četiri funkcije različite težine. Funkcije easy, medium i hard generirane su kao primjer funkcija različitih težina u kojima sigurno postoji ovisnost izlaza o ulazu dok za funkciju zraka ta povezanost nije jasna. Na samome kraju prikazani su dobiveni rezultati.Today, neural networks are an emerging topic of interest. By emulating the functions and connections of neurons in the human brain, artificial neural networks are able to learn and make predictions. While many different techniques can be used in the process of learning, this work focuses on using genetic algorithms. That technique is called neuroevolution. Two neural networks were implemented. The first one had a fixed structure where only the weights were optimized by the genetic algorithm. For the second one, genetic algorithm NEAT was used to improve the architecture. Neural networks were then used to solve regression problems by making predictions of four different functions. Easy, medium and hard functions are an example of functions where the connection between outputs and inputs is certain and clear. That is not the case with the air function. Results are shown in the final part of the work