Evaluating Non-Hierarchical Overflow Loss Systems Using Teletraffic Theory and Neural Networks

Abstract

The Information Exchange Surrogate Approximation (IESA) is a powerful tool for estimating the blocking probability of non-hierarchical overflow loss systems (NH-OLSs), but can exhibit significant approximation errors in some cases. This letter proposes a new method of evaluating the blocking probability of generic NH-OLSs by combining machine learning with IESA. Specifically, we modify IESA by using neural networks (NN) to tune a newly introduced parameter in the IESA algorithm. Extensive numerical results for a simple NH-OLS show that our new hybrid method, which we call IESA+NN, is more accurate and robust than both base IESA and direct NN-based approximation of NH-OLS blocking probability, while remaining much more computationally efficient than computer simulation. Furthermore, due to the generic nature of our technique, IESA+NN is also easily extensible to more specialized stochastic models for communications and service systems, where base IESA has previously been applied.