DeepStack is an algorithm for imperfect information games that combines recursive reasoning, decomposition, and learned intuition to handle the complexity of poker. It defeats professional poker players in heads-up no-limit Texas hold'em (HUNL) with statistical significance. Unlike previous methods that rely on abstraction, DeepStack uses a depth-limited lookahead and learned value function to approximate Nash equilibrium strategies. It avoids maintaining a complete strategy and instead computes actions on-the-fly, using a neural network to estimate the value of holding different private cards in various situations. This approach produces strategies that are more difficult to exploit than prior methods. DeepStack's performance was evaluated against professional players, showing it won 492 mbb/g, a result over 4 standard deviations from zero. The algorithm's theoretical soundness is supported by a theorem showing its exploitability is bounded by a function of the error in the value function and the number of iterations. DeepStack's approach represents a significant advancement in AI for imperfect information games, offering a new paradigm for solving large sequential games.DeepStack is an algorithm for imperfect information games that combines recursive reasoning, decomposition, and learned intuition to handle the complexity of poker. It defeats professional poker players in heads-up no-limit Texas hold'em (HUNL) with statistical significance. Unlike previous methods that rely on abstraction, DeepStack uses a depth-limited lookahead and learned value function to approximate Nash equilibrium strategies. It avoids maintaining a complete strategy and instead computes actions on-the-fly, using a neural network to estimate the value of holding different private cards in various situations. This approach produces strategies that are more difficult to exploit than prior methods. DeepStack's performance was evaluated against professional players, showing it won 492 mbb/g, a result over 4 standard deviations from zero. The algorithm's theoretical soundness is supported by a theorem showing its exploitability is bounded by a function of the error in the value function and the number of iterations. DeepStack's approach represents a significant advancement in AI for imperfect information games, offering a new paradigm for solving large sequential games.