Machine learning of irreducible error floor in the space-time trellis code

Abstract

The phenomenon of irreducible error floor in the space-time trellis code (STTC) is not fully understood. This comes from the fact that the connection between the trellis structure of the generator matrix G and the instigation of an irreducible error floor is uncertain. Given this difficulty, the present study attempts to gain a better insight into the ordeal via a data-driven approach. The classification and regression trees (CART) machine learning model is employed to predict the occurrence of the irreducible error floor from the trellis structure. Further analysis of the combinatorial characterisation of the trellis structure unveils a series of dominant patterns that consistently instigate the irreducible error floor. Furthermore, simulation also reveals that the codewords within the 'initial state' of the trellis structure are primal in the occurrence of the irreducible error floor. CART can achieve approximately 0.92 accuracy in predicting the irreducible error floor, with an average prediction time of 0.3833 μs.