Model Selection Based on Kalman Temporal Differences Learning

In this work we discuss how useful Kalman Temporal Difference (KTD) is for the purpose of improvement of multiple model learning. By KTD we mean a learning framework by combining Kalman Filters and Temporal Difference (TD) to enhance multi-agent environment. In this approach, we have to attack dependency issues against initialization parameters: the results (quality and efficiency) heavily depend on the parameters. In this investigation, we propose a new approach by estimating multiple models in parallel and by selecting suitable ones eventually.