Learning Automaton Induced Brain-Actuated Takagi-Sugeno Speed Modulation for Position Control of a Rehabilitative Robot Arm

The paper introduces a novel 2-loop architecture of brain-actuated position control system of a robot arm, realized with a Takagi-Sugeno type fuzzy speed modulator placed in the outer loop and a traditional speed-controller placed in the inner loop. The Takagi-Sugeno type speed modulator modulates the speed based on the degree of fuzziness of positional error and its time-derivative. Reinforcement Learning is employed here to autonomously determine the parameters involved in the Takagi-Sugeno type speed modulator, so as to minimize the steady-state positional error of the robot arm. The proposed Reinforcement Learning is synthesized by a Learning Automaton (matrix), whose row indices correspond to antecedents of Takagi-Sugeno fuzzy rules and columns designate possible choice of a pair of parameters. The learning in the automaton is initialized with randomly selected probabilities, so that each row-sum is unity. The automaton generates trial actions when a specific row condition is instantiated by measurement. At the end of the learning epoch, the steady-state error is measured. If it is below a user-defined threshold, the selected actions are reinforced, by increasing probabilities for the selected action, and decreasing probabilities of the non-selected actions, so that row-sum remains unity. If the steady-state error after a learning epoch is above the threshold, the selected actions are penalized. After the Learning Automaton converges, it is used for parameter selection of the Takagi-Sugeno speed-modulator based on the highest probability in each row. The complete robot arm was tested on healthy and neuro-motor patients, and the experimental results are encouraging.