Learning with assistance based on evolutionary computation

This paper proposes a learning method which learns a motion by employing an assistance in order to simplify it. This learning is done from easy to difficult level. Using genetic algorithm, it searches for the parameters of a controller appropriate for controlling the motion to be learnt by gradually increasing difficulty, i.e., by gradually decreasing the degree of assistance. We show that this gradual search enables genetic algorithm to evolve a population of controllers efficiently by giving two examples: stable riding of a bicycle and stable controlling of a double inverted pendulum. A bicycle is much easier to control when it is running at a certain velocity. An initial velocity is given as assistance and it is decreased gradually. Similarly a double inverted pendulum is much easier to control when an upward force supports the distal end of the pendulum. The reduction rate of assistance is adjustable in accordance with the adaptability of a population to the reduction