Multiple model-based reinforcement learning for nonlinear control

This paper proposes a reinforcement learning scheme using multiple prediction models (multiple model-based reinforcement learning, MMRL). MMRL prepares multiple pairs, consisting of the prediction model used to predict the future state of the control object and the reinforcement learning controller used to learn the control output. Using a soft-max function of the prediction error of each prediction model, the "responsibility signal" is calculated, which takes a larger value for the module with a more accurate prediction. By weighting the learning and the control output of each module by means of the responsibility signal, modules to deal with various situations are formed. In order to achieve a robust modular structure of MMRL without a priori knowledge, such as (lie number of modules and the region to be covered, a prior responsibility signal is formulated, assuming spatial and temporal continuity. As a method for efficient implementation of MMRL, an optimal controller (MLQC) based on multiple linear prediction and quadratic reward models is formulated. In order to verify the performance of MLQC, a simulation was performed on the swing-up of a single pendulum. It was shown that the linear prediction model and the corresponding controller were acquired by learning for the range near the suspended point and upright point of the single pendulum. The task can be learned in a shorter time than in the conventional method, and it is possible to handle redundancy of modules. (c) 2006 Wiley Periodicals, Inc.