The strategic control of an ant-based routing system using neural net q-learning agents

Abstract

Agents have been employed to improve the performance of an Ant-Based Routing System on a communications network. The Agents use Neural Net based Q-Learning approach to adapt their strategy according to conditions and learn autonomously. They are able to manipulate parameters that affect the behaviour of the Ant-System. The Ant-System is able to find the optimum routing configuration with static traffic conditions. However, under fast-changing dynamic conditions, such as congestion, the system is slow to react; due to the inertia built up by the best routes. The Agents reduce this drag by changing the speed of response of the Ant-System. For best results, the Agents must cooperate by forming an implicit society across the network.