A cascading constructive trajectory learning algorithm for block-diagonal recurrent neutral networks
Shyamala C. Sivakumar, William Robertson, William Phillips · 2002
This paper considers the construction method for the block-diagonal recurrent neural network (BDRNN) that is capable of modelling plants with complex eigenvalues. If nonlinear dynamics can be decoupled into a dominant dynamic and several less dominant dynamics, then it is feasible to employ blocks of BDRNNs, one each, to automatically learn each of these dynamics. The advantages of such an approach are that the size of the network is determined automatically and methodically, and faster learning time in comparison with a larger initial network that may be used to learn the overall dynamics. The proposed cascading constructive trajectory learning algorithm constructs a series of BDRNNs whose combined output is required to model the desired dynamic trajectory under consideration. The basic block is directly trained on the desired trajectory being learned, while, each additional cascading block is trained on the residual error between the most recent estimate and the desired trajectory.