Monte Carlo Tree Search: Random Sampling Meets Strategic Depth
Monte Carlo Tree Search made computers competitive at Go a decade before AlphaGo, by simulating thousands of random games.
For decades, game-playing AI relied on minimax search, looking ahead multiple moves and evaluating positions using hand-crafted heuristics. This worked magnificently for chess, where good heuristics (material balance, king safety, pawn structure) were well understood. In Go, it failed badly. The branching factor was too high, and nobody could write heuristics that reliably estimated position strength. Go programs in 2005 played at the level of weak amateurs.
Monte Carlo Tree Search, developed independently by Rémi Coulom and others around 2006, took a radically different approach. Instead of evaluating positions analytically, evaluate them statistically: from any position, play out thousands of random games (Monte Carlo simulations) and count the win rate. Positions that lead to more wins are better. Combine this with intelligent tree traversal, focusing simulations on promising branches, and you get MCTS: a method that requires no domain-specific heuristics at all.
The impact on Go was immediate and dramatic. Within two years, MCTS programs went from weak amateur to strong amateur level. By 2012, the best MCTS programs could challenge professional players with a handicap. This was the state of the art when DeepMind began work on AlphaGo in 2014. AlphaGo combined MCTS with deep neural networks, using neural networks to bias which nodes to explore and to estimate position value, and the combination achieved what neither alone could: superhuman Go play.
MCTS has since found applications far beyond games: planning in robotics, protein structure prediction, code generation, and drug discovery. Its core insight, that random sampling can approximate evaluations too complex to compute analytically, generalises to any decision problem with a simulatable environment. The algorithm that first made computers competitive at Go turned out to be one of the most versatile ideas in modern AI.
What it actually meant
MCTS worked without heuristics because random simulation is a lower-variance estimator than deterministic approximation given enough samples. Stochastic exploration is sometimes more reliable than clever approximation. Randomness, structured correctly, beats cleverness.
What enterprises get wrong today
Enterprises skip simulation as too expensive and use deterministic point predictions for planning in variable environments. They model customer behaviour as static, constraints as fixed, conditions as predictable. A system that does not simulate possible futures gives false confidence about expected outcomes.
Related capabilities
More from Algorithms
Ask about this story
Monte Carlo Tree Search made computers genuinely competitive at Go a decade before AlphaGo, purely by simulating random games and counting wins. No heuristics, no domain knowledge. What would you like to explore?