The author presents a learning algorithm and capabilities of perceptron-like neural networks whose outputs and inputs are directly connected to plants just like ordinary feedback controllers. This simple configuration includes the difficulty of teaching the network. In addition, it is preferable to let the network learn so that a global and arbitrary evaluation of the total responses of the plant will be optimized eventually. In order to satisfy these needs, genetic algorithms are modified to accommodate the network learning procedure. This procedure is a kind of simulated evolution process in which a group of networks gradually improves as a whole, by crossing over connection weights among them, or by mutational changes of the weights, according to fitness values assigned to each network by a global evaluation. Simulations demonstrate that these networks can be optimized in terms of various evaluations, and they can discover schemes by themselves, such as state estimation and nonlinear control.