Enhancement of extreme learning machine for estimating blocking probability of OCS networks with fixed-alternate routing

In previous work, we proposed a neural network approach to estimate the blocking probability of optical networks with fixed routing. The neural network was implemented by the extreme learning machine (ELM) framework, in which the training inputs were the optical network parameters, and the output was the overall blocking probability. The numerical results showed that the neural-network-based estimation was accurate and thousands of times faster than a computer simulation. In this paper, we apply the neural network approach to optical circuit switching (OCS) networks with fixed-alternate routing and improve the training method by using an enhancement of ELM framework. Unlike the previous ELM framework, the enhancement of ELM framework provides a random-search-based selection phase for the hidden nodes during the training step. As a result, similar performance can be achieved using fewer hidden nodes than the previous ELM framework. The numerical results show that the new enhancement of ELM training algorithm provides more accurate blocking probability estimates while reducing the required number of hidden nodes by a third compared with the previous ELM training algorithm. Furthermore, for some light traffic loading situations, our new training algorithm is hundreds of times more accurate than the existing well-known analytical approximation method.