An adaptive multi-agent routing algorithm inspired by ants behavior Gianni Di Caro and Marco Dorigo IRIDIA – Université Libre de Bruxelles – Belgium {gdicaro, mdorigo}@ulb.ac.be Abstract. This paper introduces AntNet, a novel adaptive approach to routing tables learning in connectionless communications networks. AntNet is inspired by the stigmergy communication model observed in ant colonies. We compare AntNet with the current In- ternet routing algorithm (OSPF), some old Internet routing algorithms (SPF and distrib- uted adaptive Bellman-Ford), and recently proposed forms of asynchronous online Bell- man-Ford (Q-routing and Predictive Q-routing). In all the experimental conditions con- sidered AntNet outperforms the competing algorithms, where performance is measured by standard measures such as network throughput and average packet delay, 1. Introduction Real ants are able to find shortest paths using as only information the pheromone trail deposited by other ants [1]. Ant colony optimization (ACO) algorithms which take inspiration from ants' behavior in finding shortest paths have recently been success- fully applied to combinatorial optimization [3,6,11,12,13]. In ant colony optimization a set of artificial ants collectively solve a combinatorial problem by a cooperative ef- fort. This effort is mediated by stigmergetic communication [3, 14], that is, a form of indirect communication of information on the problem structure ants collect while building solutions. In this paper we present AntNet, a novel ACO algorithm applied to the routing problem in connectionless communications networks. In AntNet artificial ants collec- tively solve the routing problem by a cooperative effort in which stigmergy plays a prominent role. Ants build local models of the network status and adaptive routing ta- bles using indirect and noncoordinated communication of information they collect while exploring the network. We compare AntNet on a variety of realistic experimental conditions with the fol- lowing state-of-the-art routing algorithms: Open Shortest Path First (OSPF) [16], Shortest Path First (SPF) [15], distributed adaptive Bellman-Ford [18], and to some recently proposed versions of asynchronous online Bellman-Ford [4, 5]. AntNet was the best performing algorithm in all considered cases. 2. Problem Characteristics and Communication Network Model Routing algorithms have the goal of directing traffic from sources to destinations maximizing some measure of network performance. Often throughput (correctly de- livered bits per time unit) and packet delay (sec) are the performance measures taken into account. Throughput measures the quantity of service that the network has been