Efficient sub-genome neuroevolution via probabilistic selection of genes
University of New Brunswick
Chimera, a novel sub-genome neuroevolution method, solves double pole balancing without velocity (DPNV), a modern version of the classic physical control machine-learning benchmark, in significantly fewer network evaluations than prior sub-genome neuroevolution methods. Neural networks are bio-mimicking computers, distributed, fault-tolerant, and applied to cashing cheques, natural language processing, and many other deep learning applications [11,16]. Deep learning repeatedly modifies a single neural network to produce known desired outputs to training example inputs. Alternatively, neuroevolution (NE) differentially reproduces candidate networks, based on their relative ability to produce training example output, or some other fitness measure . Cooperative NE methods, herein referred to as sub-genome methods, select sub-genomes, or partial networks, rather than whole genomes, or complete networks. Selecting sub-genomes is theorised to select for many independent specialisations, whereas genome selection converges on a single generalisation [5–7, 13] . Sub-genome neuroevolution has historically been made more efficient by selecting smaller more numerous genes (or specialisations). Symbiotic Adaptive Neuroevolution (SANE) randomly composes genomes from a population of chromosomes, and assigns each a fitness value. SANE then selects chromosomes by the average fitness of every genome they were a part of . Enforced Sub-populations (ESP) improves upon SANE by maintaining separate sub-populations of chromosomes for each of the x chromosomes required to form a genome . Cooperative Synapse Neuroevolution (CoSyNE) further improves upon ESP by maintaining sub-populations of genes, one for each gene in a complete genome, rather than sub-populations of multi-gene chromosomes. CoSyNE first selects genomes, but then shuffles the gene subpopulations of selected genomes [5, 6]. This thesis proposes a novel sub-genome neuroevolution method, Chimera, and compares this new method with prior methods for solving DPNV. Chimera simply selects genes from gene sub-populations, with a probability proportional to the fitness of each gene’s genome. Chimera combines CoSyNE’s gene matrix population data structure  with ESP’s intra-sub-population reproduction  and a probabilistic  variant of the fitness-proportional selection originally abandoned by SANE . Chimera uses significantly fewer network evaluations to solve DPNV than any of the prior sub-genome methods examined: SANE, ESP, or CoSyNE. While Chimera uses fewer evaluations, it can use more evaluation steps in tasks where fitness is proportional to total evaluation steps.