Exploiting correlation and budget constraints in Bayesian multi-armed bandit optimization