On the normalized maximum likelihood and Bayesian decision theory