Limited Evaluation Cooperative Co-evolutionary Differential Evolution for Large-scale Neuroevolution

  • 2018-04-19 15:34:39
  • Anil Yaman, Decebal Constantin Mocanu, Giovanni Iacca, George Fletcher, Mykola Pechenizkiy
  • 4

Abstract

Many real-world control and classification tasks involve a large number offeatures. When artificial neural networks (ANNs) are used for modeling thesetasks, the network architectures tend to be large. Neuroevolution is aneffective approach for optimizing ANNs; however, there are two bottlenecks thatmake their application challenging in case of high-dimensional networks usingdirect encoding. First, classic evolutionary algorithms tend not to scale wellfor searching large parameter spaces; second, the network evaluation over alarge number of training instances is in general time-consuming. In this work,we propose an approach called the Limited Evaluation CooperativeCo-evolutionary Differential Evolution algorithm (LECCDE) to optimizehigh-dimensional ANNs. The proposed method aims to optimize the pre-synaptic weights of eachpost-synaptic neuron in different subpopulations using a CooperativeCo-evolutionary Differential Evolution algorithm, and employs a limitedevaluation scheme where fitness evaluation is performed on a relatively smallnumber of training instances based on fitness inheritance. We test LECCDE onthree datasets with various sizes, and our results show that cooperativeco-evolution significantly improves the test error comparing to standardDifferential Evolution, while the limited evaluation scheme facilitates asignificant reduction in computing time.

 

Quick Read (beta)

loading the full paper ...