Neuroevolution: Randomness is the Simplest Thing?
Yury Tsoy · 2015
The paper describes results of the research of neuroevolutionary (NE) algorithm with random scheme for selection of operators, which modify weights and structure of evolving artificial neural networks (ANNs). Not looking at the simplicity of the algorithm some already known heuristics for improvement of efficiency of the NE algorithms are observed in result of the algorithm analysis, for example: prevalence of weights mutation over other operators; decrease of mutation probability over time; advantage of using mutation of activation functions. Interestingly that some novel heuristics are found as well: denial to use node removal operation; adaptation of connections addition/removal depending on the initial ANN size. It was also found that a limited growth of the number of nodes and connections is inherent to neuroevolutionary algorithm and is quadratic in most cases. One of the most unexpected outcomes of the research is that analysis of usage of different mutation operators lead to an observation that many non-connected problems share similar properties in the sense of application of operator A after operator B, which gives strong hopes towards creation of algorithms, which can transfer knowledge while solving different problems. On top of that the algorithm was able to solve successfully several known benchmark problems, including the Artificial Ant problem (with 20% success rate), which is quite surprising for such a simple algorithm.