This thesis is a preliminary investigation into the use of neuro-fuzzy learning techniques used to control the height, distance and angular momentum of a single running stride of a simulated linked leg. Learning from experience is the key to achieving autonomous intelligent behavior. It makes this technique more attractive than techniques that use the physical equations of motion to control the robot. Local training enables the robot to learn new strides without global off-line training.; This thesis presents the design of the controller and the results obtained in preliminary training on test data. The controller is trained only to the extent necessary to identify and analyze the trends in performance caused by changes in the various training parameter values. Also constraints are introduced to reduce the problem size. Hence this thesis is a partial solution and provides an insight into the design and appropriate training parameter values for this problem.
展开▼
