Assembler Encoding with Evolvable Operations

Abstract

Assembler Encoding is a neuro-evolutionary method which represents a neural network in the form of a linear program. The program consists of operations and data and its goal is to produce a matrix including all the information necessary to construct a network. In order for the programs to produce effective networks, evolutionary techniques are used. A genetic algorithm determines an arrangement of the operations and data in the program and parameters of the operations. Implementations of the operations do not evolve, they are defined in advance by a designer. Since operations with predefined implementations could narrow down applicability of Assembler Encoding to a restricted class of problems, the method has been modified by applying evolvable operations. To verify effectiveness of the new method, experiments on the predator-prey problem were carried out. In the experiments, the task of neural networks was to control a team of underwater-vehicles-predators whose common goal was to capture an underwater-vehicle-prey behaving by a simple deterministic strategy. The paper describes the modified method and reports the experiments.