Discover & eXplore Neural Network (DXNN) Platform, a Modular TWEANN

TL;DR

This paper introduces DXNN, a modular, hierarchical TWEANN system that enhances scalability, efficiency, and diversity in neuroevolution, outperforming existing methods on control tasks.

Contribution

The paper presents a novel modular TWEANN with hierarchical topology, new encoding, selection, and tuning methods, and a two-phase evolution approach that improves scalability and solution compactness.

Findings

DXNN outperforms state-of-the-art TWEANNs on control tasks.

The system produces highly compact solutions faster.

It effectively explores new sensors and actuators.

Abstract

In this paper I present a novel type of Topology and Weight Evolving Artificial Neural Network (TWEANN) system called Modular Discover & eXplore Neural Network (DXNN). Modular DXNN utilizes a hierarchical/modular topology which allows for highly scalable and dynamically granular systems to evolve. Among the novel features discussed in this paper is a simple and database friendly encoding for hierarchical/modular NNs, a new selection method aimed at producing highly compact and fit individuals within the population, a "Targeted Tunning" system aimed at alleviating the curse of dimensionality, and a two phase based neuroevolutionary approach which yields high population diversity and removes the need for speciation algorithms. I will discuss DXNN's mutation operators which are aimed at improving its efficiency, expandability, and capabilities through a built in feature selection method that allows for the evolved system to expand, discover, and explore new sensors and actuators. Finally I will compare DXNN platform to other state of the art TWEANNs on a control task to demonstrate its superior ability to produce highly compact solutions faster than its competitors.