Extreme learning machine for estimating blocking probability of bufferless OBS/OPS networks

Recently, the neural network approach for the blocking probability evaluation on optical networks was proposed, in which the inputs of the neural network were the optical network parameters and the output was the blocking probability of the optical network. The numerical results showed that its evaluation speed of the blocking probability was thousands of times faster than that of the discrete event simulator. However, the existing approach had two drawbacks. First, when the blocking probability was small, there was a significant approximation error due to the high dynamic range of the blocking probability. Second, the single-hidden-layer feedforward network model was used, which needed some time-consuming training algorithms to learn the parameters of hidden nodes, such as backpropagation. To solve these problems, this paper proposes to use the mean squared error of the log blocking probability as the training objective and use the extreme learning machine (ELM) framework for the training. Our numerical results show that the blocking probability estimated by our training objective is much more accurate than that of the existing approach, and it is obtained efficiently due to the greatly simplified training procedure offered by the ELM.