In this work we consider unstable control objects such as an inverted pendulum. Two evaluation procedures in genetic algorithm (GA) are set. The first involves the following steps : set two limits, -Θ and Θ, on both sides of the unstable equilibrium point, set an initial point θo in [-Θ, Θ], initiate a motion, measure the time when the motion reaches one limit, repeat simulations of neuro-control, select neural networks in order of length of holding times, and apply GA-crossover to superior neural networks of long holding times. The second involves the following steps : select neural networks in order of shortness of settling time to the equilibrium point, and apply GA-crossover to superior neural networks of short settling times. We adopt only the first evaluation procedure in the early generation stages of GA. After the number of neural networks of controllability reaches a sufficient percentage of all the neural networks in a computer, we adopt the second evaluation procedure, and GA evolution is continued. Neural networks of controllability appear at about the 10th generation and evolve to the ability limit predetermined by the structure of neural networks.