Evolving Block-based Neural Networks for Dynamic Environments

University Of Tennessee Knoxville, Knoxville TN

Investigators

Abstract

The goal of this project is to design and implement block-based neural networks (BbNNs), an evolvable neural network model suitable for dynamic environments. A BbNN consists of a 2-D array of basic neural network blocks with 4 variable input/output nodes. Each node can be either an input or an output depending on the structure represented by the signal flow. A basic block can have different internal configurations according to the number of input and output nodes. The structure and connection weights of the BbNN can be encoded as a binary string in a chromosome. Such binary representation of the BbNN enables simultaneous optimization of network structure and connections weights using evolutionary algorithms. Genetic algorithms find a near optimal structure/weight setting among many possible choices of structure and weight combinations. Chromosomes, or candidate solutions, can be directly converted to configuration bit strings of reconfigurable digital hardware as field programmable logic arrays (FPGAs) for hardware implementation of BbNNs. Unlike conventional neural network models, BbNNs contain small-sized reconfigurable neural networks as a building block. The highly regular structure of the BbNNs enables effective representation of network structure and connection weights using the signal flow. The signal flow representation enables simultaneous optimization of structure and weights through the use of evolutionary algorithms. Incorporating evolutionary optimization and modularity into neural networks will bring highly adaptive systems to dynamic environments. It will generate chips which learn themselves how to rewrite themselves for optimal performance in whatever task then are applied to. Broader impact: Evolving block-based neural networks may find applications in areas like pattern recognition, robot navigation, adaptive signal processing, control, and exploration. With their inherently modular structure, BbNNs can be easily expanded to handle large-scale problems.