An Application Research Of Reinforcement Learning In Studying Users

When a user searches the web each time, he is accessing pages pertaining to some topics. Most search engines are not sensitive to user's interests. An improved interface for the user would rank results according to the user's profile. At the same time, the information that an engine finds by the keywords a user inputs is generally far from his needs. It is more and more important whether an engine fleetly and accurately studies user's interests and knows user's recent needs for it is directly related to the credit extent a user prefers the engine. Therefore, it is more and more indis_ pensable to a user agent based an efficiently arithmetic of studying user's interests in the domain-specific search engines. Due to these basic problems and also because of the need to include information on specific domain, this paperproposes an efficient strategy of studying user's interests, which is based on Temporal Difference method, a Re_ inforcement Learning algorithm, and vector representation. The significant characteristics of the strategy are as follows: using a vector to denote user modeling and document information, from a macroscopical and microcosmic point of view analyzing the mapping relation between user model and reinforcement learning, using the implicit and explicit feedback to study user's interests, dealing with a vector v with |v|=1, and extending several algorithms related to the exact questions in this paper. At the end, we have done an experiment. From the experiment results we conclude that after using our strategy to study user's interests, final user's model is better and higher in integration, accuracy and speed to the desired goal than Young-Woo Seo in Seoul National University did.