AlphaGo and Monte Carlo tree search: The simulation optimization perspective

In March of 2016, Google DeepMind's AlphaGo, a computer Go-playing program, defeated the reigning human world champion Go player, 4-1, a feat far more impressive than previous victories by computer programs in chess (IBM's Deep Blue) and Jeopardy (IBM's Watson). The main engine behind the program combines machine learning approaches with a technique called Monte Carlo tree search. Current versions of Monte Carlo tree search used in Go-playing algorithms are based on a version developed for games that traces its roots back to the adaptive multi-stage sampling simulation optimization algorithm for estimating value functions in finite-horizon Markov decision processes (MDPs) introduced by Chang et al. (2005), which was the first use of Upper Confidence Bounds (UCBs) for Monte Carlo simulation-based solution of MDPs. We review the main ideas in UCB-based Monte Carlo tree search by connecting it to simulation optimization through the use of two simple examples: decision trees and tic-tac-toe.