Even the potential method, which seems to be the most effective method for automatic planning of the collision-free manipulator path among obstacles, uses transformation of real space coodinates of the obstacle to joint angle space, so the analyses become difficult or impossible when the manipulator degrees of freedom become large. In our previous study, we decreased this difficulty through the use of fuzzy reasoning which also enabled the consideration of much useful different dimensional information, for example, the deviation from the target, the proximity to the obstacles and the rate of change of these quantities. However, an effective membership function or rule for good efficient planning is as yet unestablished. This is a general problem in fuzzy reasoning. Furthermore, the size of the lattice gap has a large effect on search efficiency, which is poor when the target point and obstacles are far from the manipulator under small and uniform lattice gaps. This paper proposes a plan to improve the above two points. We applied a Boltzmann machine of neural network to optimize the membership function, and the size of lattice gaps was made flexible according to the magnitude of deviation and proximity using fuzzy reasoning. As a result, planning in all ranges of manipulator action was achieved, and the search efficiency was better than in the previous study.