An adaptive backpropagation through time training algorithm for a neutral controller

An adaptive method is proposed for choosing the time length (or number of time steps) used in the standard backpropagation through time (BTT) algorithm. BTT is commonly used in training neurocontrollers to perform control actions for a given cost function. Traditionally, the time length of BTT is chosen based on experience, and sometimes unnecessary calculations are performed because the time length is too large or too small. The adaptive backpropagation through time (ABTT) algorithm proposed is a technique for finding a minimum time length required to train a controller successfully to achieve its objective using the BTT algorithm. A DC motor controller is used as an example to demonstrate its feasibility. The performance of adaptive and standard BTT is compared in terms of objective achievement, cost function and computation requirements.<>