An MDP-based peer-to-peer search server network

A distributed search system consists of a large number of autonomous search servers logically connected in a peer-to-peer network. Each search server maintains a local index of a collection of documents available at the server or on other peer machines. When a query is received by any server in the network, a distributed search process determines the most relevant search servers and redirects the query to them for processing. We model the distributed search process as Markov decision processes (MDPs). The estimated relevance of a server to a query is regarded as the reward in the MDP model. Once the MDP policies representing the global knowledge are obtained at each server through asynchronous value iteration, the most relevant servers to a given query can be efficiently identified despite the lack of centralized control and global knowledge at each autonomous server. We discuss the implementation and complexity of the asynchronous value iteration and how we extend the traditional MDP to handle the multiple-access policy (i.e., more than one optimal server is returned) and queries with multiple terms. Finally, experiments are conducted using the TREC collection. We show that the MDP-based distributed search can achieve results very close to that of a centralized search.