Using implicit fitness functions for genetic algorithm-based agent scheduling

In a distributed network, servers have various capabilities that make them more suitable for certain tasks. Mobile agents move from server to server based on various scheduling algorithms. Choosing a server for a particular agent can be difficult for many users. We present the design of a system which uses a genetic algorithm (GA) to “learn” the type of server best suited to an agent based on parameters such as bandwidth, latency, and CPU availability supplied by the Network Weather Service (NWS). Unlike typical GA implementations, we use an implicit fitness function defined as the amount of work done per unit time on the actual servers. We also use a ‘hinting’ system to further improve results through allowing agents to share information on server performance for particular agents. We provide experimental results indicating that the GA performance improves significantly over time, and provides a significant advantage over round-robin scheduling.