Abstract
Over the last two decades, peer-to-peer systems have proven their vital role in sharing various resources and services to diverse user communities over the internet. The unstructured P2P network is the most popular topology, and the resources are fully distributed among participating peers. Therefore, searching is a challenging issue due to the absence of control over resource locations. Intelligent decisions should be made to select a particular number of neighbors that can hold relevant resources for queries instead of selecting neighbors randomly. In this paper, an intelligent neighbor selection (INS) algorithm is introduced that uses a reinforcement learning approach, 'Q-learning'. The main objective of this algorithm is to achieve better retrieval effectiveness with reduced searching costs by fewer connected peers, exchanged messages, and less time. To achieve this, INS relies on Q-learning, which makes a Q-table in each peer and stores the Q-values gathered from the results of previously sent queries, and uses them for the forthcoming queries. The cold start issue during the training phase is also addressed in this research, which allows INS to improve its results continuously. The simulation results show a significant improvement in searching for a resource with compression to controlled flooding and learning processes after sufficient training. Here, retrieval effectiveness, search cost in terms of connected peers, and average overhead are 1.23, 104, 167, respectively.