Self-organization of evaluation space for planning using state discretization resolution

Hierarchical structures of multiple modules are investigated to solve large and complex problems. Many studies have effectively utilized the abstraction of successive states, but there have been no sufficiently detailed discussions of the overlap between these abstracted states and continuity between multiple modules. Here, we propose organization of the evaluation space considering the continuity of multiple rewards. Multiple reward functions are applied to the planning scheme and each reward function provides different discretization of the state space. Based on the use of multiple resolutions, we propose integration of the evaluation space. The objective of this integration was to reduce the number of discrete state sets used to calculate the value function of the task. We report here an example of effective use of multiple resolutions based on multiple rewards is effective for planning in a two-DOF manipulator motion planning problem with redundant tactile sensors.