A New Q-learning with Generalized Approximation Spaces

For measuring the uncertainty of behavior, the average rough coverage doesn't consider the difference among middle learning stages in reinforcement learning. To address this problem, a novel measure model based on generalized approximation spaces is proposed. In this study, uncertainty is regarded as the local feature of a state and used to guide future learning. Data-driven Q-learning based this novel model is presented for improvement of strategies based exploration. The measure function of uncertainty is used to control the balance between exploration and exploitation. Experiment results show that data-driven reinforcement learning is effective.