A behavior learning algorithm for unmanned underwater vehicles

This paper proposes the behavior learning algorithms for improving the performance of unmanned underwater vehicles (UUVs). Basically, the motion of a UUV with behavior-based controls is determined by the behaviors' outputs which are calculated based on the uncertain dynamic models. Therefore, the performance of a UUV can be improved by learning the dynamic models. For this purpose, the perturbation models for three behaviors such as speed command, turning motion, and diving motion are derived and trained by supervised neural network so that the perturbations are estimated and compensated in the exploitation stage. Simulation result is presented for verifying the performance of the learning algorithms.