Evolving Neural Networks through Augmenting Topologies

Abstract

An important question in neuroevolution is how to gain an advantage from evolving neural network topologies along with weights. We present a method, NeuroEvolution of Augmenting Topologies (NEAT), which outperforms the best fixed-topology method on a challenging benchmark reinforcement learning task. We claim that the increased efficiency is due to (1) employing a principled method of crossover of different topologies, (2) protecting structural innovation using speciation, and (3) incrementally growing from minimal structure. We test this claim through a series of ablation studies that demonstrate that each component is necessary to the system as a whole and to each other. What results is signicantly faster learning. NEAT is also an important contribution to GAs because it shows how it is possible for evolution to both optimize and complexify solutions simultaneously, offering the possibility of evolving increasingly complex solutions over generations, and strengthening the analogy with biological evolution.

Training artificial neural networks directly on the concordance index for censored data using genetic algorithms

“…One idea for improvement would be to evolve the structure (e.g. complexity) of the ANNs together with the weights as in the NEAT algorithm [25].…”

Evolving neural fields for problems with large input and output spaces

“…The difficulties of the different versions of these tasks are well known from previous studies. Early neuroevolution methods, as well as several other machine learning methods, were not able to solve the more difficult varieties of these tasks (Stanley and Miikkulainen, 2002). Pole balancing here serves as a simple substitute for higher dimensional problems that will also require such precise tuning of connection weights.…”

“…The NEAT method (Stanley and Miikkulainen, 2002;) is a well known and competitive neuroevolution method that introduces a number of ideas to successfully deal with the problems discussed in the previous section. One idea is to give genes an unchanging identity.…”

“…There is also a more difficult version of DPNV ("Anti-wiggling" DPNV) with a different fitness function, where wiggling of the poles is punished and generalization to at least 200 out of 625 starting angles is required. These tasks have been described in more detail previously (Wieland, 1991;Stanley and Miikkulainen, 2002). While indirect encodings like HyperNEAT have obviously not been designed for these kinds of problems, which have small input and output spaces, and require few neurons and connections to be solved, pole balancing tasks certainly require fine tuning of connection weights.…”

“…We believe that these operators are also helpful for evolving large neural networks. Second, complexification from small structures, i.e., gradual increase of the network sizes during the course of evolution, has been shown to be an important reason for the success of the NEAT method (Stanley and Miikkulainen, 2002). Therefore, exploration of the search space by complexification is also used as a strategy by NEATfields.…”