Pseudo-random number generation using LSTMs

Previous studies have developed pseudo-random number generators, where a pseudo-random number is not perfectly random but is practically useful. In this paper, we propose a new system for pseudo-random number generation. Recurrent neural networks with long short-term memory units are used to mimic the appearance of a given sequence of irrational number (e.g., pi), and these are intended to generate pseudo-random numbers in an iterative manner. We design algorithms to ensure that the output sequence contains no repetition or pattern. Through experimental results, we can observe the potential of the proposed system in terms of its randomness and stability. As this system can be used for parameter approximation in machine learning techniques, we believe that it will contribute to various industrial fields such as traffic management and frameworks for sensor networks.