Distributed Biobjective Ant Colony Algorithm for Low Cost Overlay Network Routing Benjamin McBride Kansas State University 2061 Rathbone Hall Manhattan, KS 66506 Email: bdm7935@ksu.edu Caterina Scoglio Kansas State University 2061 Rathbone Hall Manhattan, KS 66506 Email: caterina@ksu.edu Sanjoy Das Kansas State University 2061 Rathbone Hall Manhattan, KS 66506 Email: sdas@ksu.edu Abstract— In this paper we introduce a biobjective ant colony algorithm for constructing low cost overlay routing networks. The ant colony algorithm is distributed and adaptive in finding shortest paths from source to destination nodes while also constructing a low cost overlay routing network. Additionally, we define a cost model for overlay network construction that includes network traffic demands. The proposed ant colony algorithm was applied to a randomly generated 100-node network with an average node degree of 10.2. The results show that the algorithm quickly converges to the shortest path between nodes while converging on a low cost overlay routing network topology, despite changing traffic demands. I. I NTRODUCTION Ant colony algorithms get there inspiration from the be- havior of real ants foraging for food. Individual ants act as simple agents that forage for food by following pheromone trails and depositing pheromone along the path taken. An ant is more likely to follow a trail with a high concentration of pheromone. Pheromone on shorter paths gets increasingly reinforced as more ants follow the higher concentration of pheromones. This is due to more ants being able to travel a shorter path than a longer path over a given period of time. Eventually, the pheromone concentration is greatest along the shortest path to the food source. This results in the ants converging on the shortest path to the food source. While each ant acts independently, the collective behavior of the ant colony is cooperative. This emergent behavior has inspired a host of stochastic optimization algorithms called ant colony algorithms. Ant colony algorithms, as described in [1], [2], and [3], have been quite popular for a wide variety of discrete opti- mization problems such as the traveling salesman problem, quadratic assignment problem, job-shop scheduling, vehicle routing, and graph coloring. This is widely due to the ease of implementation and the inherent balance between exploration and exploitation found in the ant colony algorithm. The adaptive multi-agent characteristics of ant colony algorithms are also attractive for distributed and dynamic network routing problems. Ant colony algorithms have previously been used in connection-oriented network routing, connectionless network routing, and wireless sensor network routing [4] [5] [6] [7] [8]. The growth of peer-to-peer and other multimedia applica- tions such as video conferencing and internet telephony has created the need for increased quality of service over the physical network. Providing the required quality of service and performance requirements of these applications over a packet switching network has been elusive. In recent years, overlay routing networks have been seen as a possible solution [9] [10] [11] to the quality of service problem. An overlay routing network is an application-layer, logical network created on top of the physical, or underlay network. The overlay routing network can be used to improve performance and provide quality of service by routing data through intermediate nodes in the logical network, rather than directly through the underlay network. In this manner, an over- lay routing network can respond to changing traffic congestion and performance degradation in the underlay network. In this work, overlay routing networks are created using an ant colony algorithm inspired by the Ant Colony System algorithm and the AntNet algorithm described in [2] and [5]. As data is routed through the network the algorithm converges on the shortest path between source and destination nodes, while also constructing a low cost overlay routing network. This biobjective ant colony algorithm is both distributed and adaptive. While each individual node is selfish in the con- struction of a low cost overlay, the emergent behavior of the ant colony system results in a cooperative system. This differs from other selfishly constructed networks [12]. The proposed algorithm was tested on a randomly generated network with changing traffic demand patterns. It was found that the routing quickly converges to the shortest path between source and destination nodes while also constructing a low cost overlay routing network. As the traffic demands reach equilibrium the cost of the network converges on a single overlay topology. The paper is organized as follows. In Section II we discuss background information related to the ant colony algorithm and the specific overlay cost model used. Section III gives the detailed algorithms implemented. Simulation results are pre- sented in Section IV. Concluding remarks and future research directions are outlined in Section V.