Learning method for hierarchical behavior controller

Complex behavior is difficult to obtain using an unsupervised leaning method because of the enormous search space required. In this paper, we propose the hierarchical behavior controller which consists of three types of modules: behavior coordinator, behavior controller and feedback controller. We also propose a new learning algorithm for the behavior coordinator and the behavior controller that consists of some sub-coordinators and some sub-controllers, respectively. This algorithm selects a deficient one by evaluating each sub-coordinator or sub-controller using multiple regression analysis based on previously obtained evaluation values. This can reduce the search area and the learning times by avoiding the necessity of trying to tune good sub-coordinators or sub-controllers. The hierarchical behavior controller is applied to the problem of controlling a seven-link brachiation robot, which moves dynamically from branch to branch like gibbon swinging its body.