A Q-learning approach to the continuous control problem of robot inverted pendulum balancing