A novel technique to self-adapt parameters in parallel/distributed genetic programming

  • PDF / 438,758 Bytes
  • 10 Pages / 595.276 x 790.866 pts Page_size
  • 41 Downloads / 153 Views

DOWNLOAD

REPORT


METHODOLOGIES AND APPLICATION

A novel technique to self-adapt parameters in parallel/distributed genetic programming Marco Russo1

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract This paper introduces the Supervisor Evolutionary Algorithm, a novel technique that allows for self-adapt almost all the internal parameters in parallel distributed client-server genetic programming. This novel adapting mechanism, is itself of an evolutionary nature, so we have a double evolutionary tool. The upper level, as is usual in evolutionary computing, has its own customized selection, crossover, and mutation mechanisms. The lower stage used here is the Brain Project a paralleldistributed software tool for formal modelling of numerical data using a hybrid neural-genetic programming technique. As demonstrated by the experiment reported in this paper, our approach works well adapting continuously its internal parameters. Keywords Genetic programming · Neural networks · Evolutionary computing · Parallel computing · Distributed computing

Acronyms BIs BNGPA BP CPC CPU EC GP GWC IP LT MIMO PC SAPC SEA SIs SPC TCP

1 Introduction Base Individuals Base Neuro-GP Algorithm Brain Project Client Personal Computer Central Processing Unit Evolutionary Computing Genetic Programming Game Winning Criterion Internet Protocol Learning Task Multi-Input-Multi-Output Personal Computer Self-Adaptive Parameter Control Supervised Evolutionary Algorithm Supervisor Individuals Server Personal Computer Transmission Control Protocol

Communicated by V. Loia. Project web page: superpippo.ct.infn.it/~marco.

B 1

Marco Russo [email protected] University of Catania, Catania, Italy

Evolutionary Computation (EC) is a wide and extremely heterogeneous set of algorithms for global optimization inspired by biological evolution that has been successfully applied in many application fields (Dasgupta and Michalewicz 2013; Paulinas and Ušinskas 2007; Cordón et al. 2003; De La Fraga and Coello 2011; Russo et al. 2014). More or less all the algorithms to be found in EC depend on several parameters that greatly influence the performance of the algorithm itself. In some cases a bad set of parameters can lead to very poor solutions or, even worse, to no convergence. There are two different methodologies for setting parameter values: parameter tuning and parameter control (Karafotias et al. 2015). In the first case, the set of values for the parameters is fixed before the algorithm itself starts. These parameters remain unchanged during the run. For this reason we often refer to this first method as the stationary side. In the case of parameter control the values of the parameters can change during the run. Several techniques are possible to execute this job, but one of the most interesting is that called Self-Adaptive Parameter Control (SAPC) (Brest et al. 2006; Karafotias et al. 2015; Santamaría et al. 2013; Kramer 2010; Aleti and Moser 2016; Ginley et al. 2011), the basic idea of which is often a double evolutionary approach where there is an ev