Abstract: A new training method was proposed to solve the problem that some gradients of the objects are not easy to calculate. This method was based on the principle of gravity optimizes parameters. Random initial parameter based on step was set as coordinate form which in the midpoint of the multidimensional space. The error between the actual output and the target output was set as radius. This method had advantages which could not need to calculate the gradient and could randomly select initial value. This method was successfully used in the PID controller， LQR controller and neural network.

Key words: gravity, training, optimization, gradient, neural network, PID controller, LQR controller, parameter