Towards reduction of autocorrelation in HMC by machine learning