DES-HyperNEAT: Towards Multiple Substrate Deep ANNs
Amund Tenstad, Pauline Catriona Haddow · 2021
Neuroevolution (NE), inspired by the natural evolution of biological brains, may be applied to the evolution of both weights and the topology of Artificial Neural Networks (ANNs). DES-HyperNEAT, proposed herein, further extends existing NE algorithms to enable evolvable substrate topologies with non-fixed node positions - a desirable feature for optimisation of deep ANNs. DES-HyperNEAT builds on HyperNEAT and ES-HyperNEAT whilst adding beneficial features from NEAT and MSS-HyperNEAT. The preliminary experiments in this work consider potential variants of DES-HyperNEAT, evaluating such variants in terms of performance and efficiency on the Iris, Wine and Retina datasets. Further, DES-HyperNEAT, ES-HyperNEAT, HyperNEAT and NEAT are compared to highlight whether the DES-HyperNEAT extension has merit in terms of performance and efficiency.