Intelligent Packets For Dynamic Network Routing Using Distributed Genetic Algorithm

A distributed GA is designed for the packet switched network routing problem under minimal information. The requirements of such a problem mean that agents are required to possess more intelligence than was previously the case. To this end a distributed GA approach is developed and benchmarked against the AntNet algorithm under the same information constraints. 1 GENERAL FORMATTING INSTRUCTIONS Network information systems and telecommunication in general rely on a combination of routing strategies and protocols to ensure that information sent by a user is actually received at the desired remote location. In addition, the distributed nature of the problem means that multiple users can make requests simultaneously. This results in delayed response times, lost information or other reductions to the quality of service objectives on which users judge network service. Routing is the process used to determine how a packet travels from source to destination. Protocols are used to implement handshaking activities such as error checking and receiver acknowledgements. In this work, we are interested in the routing problem on computer networks. The routing problem has several properties, which make it particularly challenging. The problem is distributed in nature; hence a solution that assumes access to any form of global information is not desirable. The problem is also dynamic; hence a solution that is sufficient for presently experienced network conditions may well be inefficient under other loads experienced by the network. Moreover, the traffic experienced by networks is subject to widely varying load conditions, making ‘typical’ network conditions unrepresentative. Traditionally, routing strategies are implemented through the information contained in routing tables available at each node in the network (Forouzan, 2001). That is, a table details the next ‘hop’ a packet takes based on the overall destination of the packet. This should not be taken to imply that a routing table consists of an exhaustive list of all destinations – a form of global information. Instead, the table consists of specific entries for the neighboring nodes and then a series of default paths for packets with any other destination – such as OSPF or BGP4 (Halabi, 1997). Application of a classical optimization technique to such a problem might take the form of first assessing the overall pattern of network traffic, and then defining the contents of each routing table such that congestion is minimized. This approach does not generally work in practice as it simply costs too much to collect the information centrally on a regular basis, where regular updating is necessary in order to satisfy the dynamic nature of network utilization. We, therefore, see the generic objectives of a routing strategy to be both dynamically reconfigurable and be based on locally available information, whilst also satisfying the user quality of service objectives (i.e. a global objective). Several approaches have been proposed for addressing these objectives including: active networking (Tennenhouse et. al.,1997), social insect metaphors (Di Caro, Dorigo, 1998), (Heusse et al., 1998) cognitive packet networks (Gelenbe et. al.,1999), and what might be loosely called other ‘adaptive’ techniques (Corne et. al., 2000). The latter typically involve using evolutionary or neural techniques to produce a ‘routing controller’ as opposed to a ‘routing table’ at each node, where the controller may require knowledge of the global connectivity to ensure a valid route. The global information assumption may be avoided by framing the problem as a reinforcement-learning context (Boyan, Littman, 1994). However, the Q-learning method, on which this is based, results in single path solutions for each destination. Both the social insect metaphor and the cognitive packet approach provide a methodology for routing, without such constraints; by utilizing probabilistic routing tables and letting the packets themselves investigate and report network topology and performance. AAAA, Alife, Adaptive Behavior, Agents, and Ant Colony Optimization All methods as currently implemented, however, suffer from one drawback or another. Cognitive packet networks and active networking algorithms attempt to provide routing programs at the packet level, hence achieving scalable run time efficiency becomes an issue. To date, implementations of ‘adaptive’ techniques and social insect metaphors have relied, at some point, on the availability of global information (Liang, et al., 2002). The purpose of this work is to investigate the application of a genetic algorithm (GA) to build on lessons learnt from the social insect metaphor. This represents a major departure from previous works attempting to utilize GAs to solve the dynamic routing problem e.g. (Corne D.W., et al. 2000). In particular, a distributed GA is defined in which populations associated with each node of the network are required to co-evolve to solve the problem as a whole. Moreover, the GA interaction with the environment drives the features measured by the routing tables, as opposed to the tables predefining the features for measurement (a form of a priori information). Section 2 introduces the ‘ant’ based social insect metaphor against which the proposed approach is compared. Section 3 introduces the proposed GA-agent scheme. Section 4 summarizes the network on which experiments are performed. Results are presented in section 5 and conclusions drawn in section 6. 2 ROUTING USING A SOCIAL INSECT METAPHORE As indicated above, active networking (Tennenhouse et. al.,1997) and cognitive packet (Gelenbe et. al.,1999) based approaches emphasize a per packet mechanism for routing. The aforementioned ‘adaptive’ techniques (Corne et. al., 2000) tend to emphasize adding ‘intelligence’ to the routers leaving the packets unchanged. A social insect metaphor provides a middle ground in which the concepts of a routing table and data packet still exist, but in addition, intelligent packets – ants – are introduced that interact to keep the contents of the routing tables up to date. To do so, the operation of ant packets is modeled on observations made regarding the manner in which worker ants use chemical trails as a method of indirect stigmergic communication. Specifically, ants are only capable of simple stochastic decisions influenced by the availability of previously laid stigmergic trails. The chemical denoting a stigmergic trail is subject to decay over time, and reinforcement proportional to the number of ants taking the same path. Trail building is naturally a bidirectional process, ants need to reach the food (destination) and make a successful return path, in order to significantly reinforce a stigmergic trail (Forward only routing has also been demonstrated (Heusse et al., 1998)). Moreover, the faster the route, then the earlier the trail is reinforced. An ant on encountering multiple stigmergic trails will probabilistically choose the route with greatest stigmergic reinforcement. Naturally, this will correspond to the ‘fastest’ route to the food (destination). The probabilistic nature of the decision, however, means that ants are still able to investigate routes with a lower stigmergic trial. This approach has proved to be a flexible framework for solving a range of problems including the traveling sales man problem (Dorigo et al., 1996) and the quadratic assignment problem (Maniezzo et al., 1999). The work reported here follows the ‘AntNet’ algorithm of Di Caro and Dorigo (Di Caro, Dorigo, 1998), and is informally summarized as follows, • Each node in the network retains a record of packet destinations as seen on data packets passing through that node. This is used to periodically, but asynchronously, launch ‘forward’ ants with destinations stochastically sampled from the collected set of destinations; • Once launched, a forward ant uses the routing table information to make probabilistic decisions regarding the next hop to take at each node. While moving, each forward ant collects time stamp and node identifier information where this is later used to update the routing tables along the path followed; • If a forward ant re-encounters a node previously visited before reaching the destination, it is killed; • On successfully reaching the destination node, total trip time is estimated and the forward ant converted into a backward ant; • The backward ant returns to the source using exactly the same route as recorded by the forward ant. Instead of using the data packet queues, however, the backward ant uses a priority queue; • At each node visited by the backward ant the corresponding routing table entries are updated to reflect the relative performance of the path; • When the backward ant reaches the source, it ‘dies’. Although providing for a robust ant routing algorithm under simulation conditions, an assumption is made, which inadvertently implies the use of global information knowledge of the number of nodes in the network (Di Caro, Dorigo, 1998). The definition of routing tables is, such that it is assumed that every node has a unique location in the routing table, see Table 1, or a total of L (number of neighboring nodes) by K (number of nodes in the entire network) entries. In practice, this is never the case. To do so would assume that it is first feasible, and secondly, should the network configuration ever change, then all nodes should be updated with the new configuration information. Moreover, as forward ants propagate across the network, the amount of information they need to ‘carry’ also increases (node identifier and time stamp). AAAA, Alife, Adaptive Behavior, Agents, and Ant Colony Optimization     î    − + − − +     = ) ( ) ( inf inf sup inf sup 2 1 I t I I I I c t W c r ant ant best Table-1 Original Routing Table at any Network Node k on the NTTnet All Network Nodes (Possible Destinations) P1,1 P1,2 -----P1,55 P2,1 P2,2 -----P2,55 --------------------O ut go in g L in ks