Bayesian Optimization in AlphaGo improves self-play win-rate from 50% to 66.5%