Topology management in unstructured P2P networks using neural networks

Resource discovery is an essential problem in peer-to-peer networks since there is no centralized index in which to look for information about resources. In a pure P2P network peers act as servers and clients at the same time and in the Gnutella network for example, peers know only their neighbors. In addition to developing different kinds of resource discovery algorithms, one approach is to study the different topologies or structures of the P2P network. In many cases topology management is based on either technical characteristics of the peers or their interests based on the previous resource queries. In this paper, we propose a topology management algorithm which does not predetermine favorable values of the characteristics of the peers. The decision whether to connect to a certain peer is done by a neural network, which is trained with an evolutionary algorithm. Characteristics, which are to be taken into account, can be determined by the inputs of the neural network.