摘要 |
PURPOSE: To learn many-valued mapping to movable state information on each movable part from position information on one end of a controlled system while reducing the calculation quantity and to control the motion of the controlled object in real time as much as possible by learning by specific many- valued function approximation. CONSTITUTION: A two-valued function approximation network 9 learnt from position direction control information (p, q,ϕ) on the tip of a robot arm which is given from the outside to compute mapping to an angle (θ0 ,θ1 ,θ2 ) of each rotary articulation. A value which is closer to a last output between two obtained values is selected and outputted. Here, one of the two values is only selected optionally as a 1st output value. The state of each articulation is controlled on the basis of the output result and the arm tip moves in a desired direction to a desired space position. This control is reduced in the quantity of calculation by the learnt two-valued function approximation network 9, so real-time processing becomes possible.
|