Reinforcement Learning-Aided Markov Chain Monte Carlo For Lattice Gaussian Sampling

Sampling from the lattice Gaussian distribution has emerged as a key problem in coding, decoding and cryptography. In this paper, the Gibbs sampling from Markov chain Monte Carlo (MCMC) methods is investigated for lattice Gaussian sampling. Firstly, the error function of random scan Gibbs sampling is derived, and we show that it is partially determined by the selection probabilities over the sampling components. Then, in order to minimize the error function for a better sampling performance, a reinforcement learning mechanism is proposed for random scan Gibbs sampling to adaptively update the selection probabilities by learning from the random samples generated along with the chain. Finally, simulation results based on MIMO detection are presented to confirm the performance gain at the expense of limited complexity cost.