Machine Learning Based End-to-End Constellation Training for Communication Systems

The constellation design is critical to the performance of communication systems. Conventional constellation schemes usually do not achieve optimal performance. Their design is usually based on some assumptions such as channel models and equal information bit probability distributions. Therefore, the performance degrades if the assumptions are unreal or oversimplified. Furthermore, effective and efficient methods to generate massive order constellations are still missing. In this paper, we propose a machine learning based constellation training method for communication systems. We use the autoencoder to train the constellation points. The proposed method trains the constellation by transmitted and received information bits. It does not depend on the assumptions such as specific channel models or information bit probability distributions. The proposed method can effectively optimize the constellation points, even for massive order constellations. Simulation results show that the proposed method outperforms conventional methods. The performance improvement is more significant when the number of bits per symbol is larger.