This paper presents an agenda-based user simulator which has been extended to be trainable on real data with the aim of more closely modelling the complex rational behaviour exhibited by real users. The trainable part is formed by a set of \emph{random decision points} that may be encountered during the process of receiving a system act and responding with a user act. A sample-based method is presented for using real user data to estimate the parameters that control these decisions. Evaluation results are given both in terms of statistics of generated user behaviour and the quality of policies trained with different simulators. Compared to a handcrafted simulator, the trained system provides a much better fit to corpus data and evaluations suggest that this better fit should result in improved dialogue performance.