Personalizing user experience with neural fitted q iteration

Abstract

Machine learning techniques can be implemented to provide personalized websites. In an example, a group of training inputs is determined. A training input of the group of training inputs can comprise a current state that indicates a visit to a website via a user device, an action that indicates a recommendation rendered via the user device during the visit to the website, a user reward that indicates whether user input associated with a user identity via the user device interacted with the recommendation, and a next state that indicates a state of the website for the user device in response to the action being taken from the current state. The training inputs can be used for performing batch offline neural fitted Q iteration training on a neural network to produce a trained neural network. The trained neural network can be stored.