Learning of fuzzy control rules can be considered as a constrained nonlinear optimization problem, in which the objective function isn't differentiable.
alpha constrained Powell's method was proposed to solve the problem.
But by the method, we must prepare the initial rules that are written by human to obtain the precise control rules.
Also, the method needs a lot of control experiments to evaluate the objective function.
We propose a new method, alpha constrained simplex method, which adopts alpha level comparison in Simplex method.
alpha level comparison compares not only the value of objective function but also the satisfaction level of constraints.
We show that we can obtain the precise control rules from mechanically created rules by our method with fewer control experiments.