Playground: A Novel

The characters in Playground obsess over an ancient board game called Go. They consider this game far too complex to ever be understood by a computer, so the defeat of the best human player by an AI player is a landmark moment in these characters’ lives. Go originated in China more than 2,500 years ago. It is played on a 19x19 grid, where players take turns placing black or white stones with the objective of surrounding more territory than their opponent. While the rules of Go are simple, the game is known for its strategic complexity and depth. The number of possible legal board configurations in Go far exceeds that of chess, making it a particularly difficult game for computers to master. Traditional brute-force methods used by earlier AI systems in chess were inadequate for Go due to its vast search space, which requires more nuanced decision-making and pattern recognition than what most previous AI techniques could handle.

For many years, Go was considered one of the final frontiers for AI research in games. Despite progress in computer science and AI, many experts believed it would take decades before AI could defeat a world-class Go player because of the game’s unpredictability and reliance on intuition and creativity, attributes traditionally associated with human cognition. AlphaGo, developed by the London-based AI company DeepMind (which Google acquired in 2014), revolutionized the approach to mastering Go. Rather than relying on brute-force computation to evaluate every possible move, as early chess programs like IBM’s Deep Blue did, AlphaGo combined advanced techniques in machine learning and neural networks to develop a much more sophisticated form of AI.

AlphaGo’s core innovation was its use of deep reinforcement learning, a branch of AI in which systems learn by interacting with environments and optimizing their behavior through trial and error. The system trained by playing millions of games against itself, using a process called supervised learning to learn from human professional games and unsupervised learning to develop its own strategies. AlphaGo used two types of neural networks: the “policy network,” which selected moves based on the current board state, and the “value network,” which predicted the probability of winning from any given position. The combination of these networks enabled AlphaGo to prune the vast tree of possible moves in Go, allowing it to focus only on the most promising ones.

Before the famous match against Go master Lee Sedol, AlphaGo had already made headlines by defeating Fan Hui, the European Go champion, in October 2015. Its 5-0 victory over Fan Hui was a surprise, but the ultimate test was the match against Lee Sedol, one of the world’s strongest Go players. In March 2016, AlphaGo and Lee Sedol faced off in a highly anticipated five-game match in Seoul, South Korea. AlphaGo won the first three games, securing its overall victory, before Lee Sedol won the fourth game, preventing a clean sweep. AlphaGo then won the final game, taking the series 4-1. The outcome shocked the Go community and AI researchers alike, who had expected Lee Sedol to win or at least put up a far stronger fight.

The most talked-about moment came in the second game when AlphaGo played a move that defied traditional Go strategies. Move 37 was widely considered unconventional and surprising, even to the human experts commentating on the game. The move was initially seen as a mistake, but it later became clear that the move demonstrated AlphaGo’s ability to think beyond the realm of human Go expertise, highlighting its creative and highly strategic approach to the game. This was the first time many realized that AI could not only replicate human play but also transcend it in ways that humans hadn’t anticipated. AlphaGo’s victory demonstrated the potential of deep learning and neural networks.