Quick acquisition of a control algorithm for small flying robots by a learning system combining an artificial neural network and an immunity algorithm

Flying robots: small computer controlled airplanes are useful for safe and economical investigation with aerial photography. We have developed such planes using DSP boards equipped with sensors and many PID control loops. However, the tuning of their control parameters is difficult because of many state variables, interference between control loops, and the nonlinear characteristics of aviation dynamics. In this paper, the self-organization of their control algorithm is proposed. An artificial neural network is used as the controller. The control algorithm is organized by back propagation type learning to decrease the difference between the desired state variables and the estimated ones. A problem of back propagation-type learning is apt to appear in local optimum solutions. Therefore, many sets of initial values of input gains at the synapses of neurons are prepared similar to DNA codes in immune bodies for immunity process simulation. When a local optimum solution of the neural network is found, the DNA code of teamed gains is memorized at the T cell of the immunity process. Cells near these having the memorized DNA are eliminated, and new cells far from them are added to the process in searching for better solutions. It is verified by computer simulation that a better control algorithm is automatically organized for small flying robots having complicated dynamics.