Enhancing Multiagent Genetic Network Programming Performance Using Search Space Reduction

TL;DR

This paper proposes a method to reduce the search space in Situation-Based Genetic Network Programming (SBGNP) by applying simplified operators, leading to improved performance on the Tileworld benchmark.

Contribution

It introduces a novel approach to apply search space reduction techniques to SBGNP, which was not previously explored.

Findings

Improved average fitness on Tileworld benchmark

Effective reduction of search space in SBGNP

Enhanced agent control performance

Abstract

Genetic Network Programming (GNP) is an evolutionary algorithm that extends Genetic Programming (GP). It is typically used in agent control problems. In contrast to GP, which employs a tree structure, GNP utilizes a directed graph structure. During the evolutionary process, the connections between nodes change to discover the optimal strategy. Due to the large number of node connections, GNP has a large search space, making it challenging to identify an appropriate graph structure. One way to reduce this search space is by utilizing simplified operators that restrict the changeable node connections to those participating in the fitness function. However, this method has not been applied to GNP structures that use separate graphs for each agent, such as situation-based GNP (SBGNP). This paper proposes a method to apply simplified operators to SBGNP. To evaluate the performance of this method, we tested it on the Tileworld benchmark, where the algorithm demonstrated improvements in average fitness.