Reinforcement learning pioneers win Turing Award for AI breakthroughs

Reinforcement learning pioneers Barto and Sutton receive computing’s highest honor for developing computational intelligence theories that powered AI breakthroughs like AlphaZero. Their work, which focuses on how computers learn through rewards and punishment, has revolutionized AI systems’ ability to master complex tasks and highlights the important role of exploration, curiosity, and play in developing artificial intelligence.

The big picture: The Association for Computing Machinery awarded the 2025 Turing Award—computing’s equivalent to the Nobel Prize—to professors Andrew G. Barto and Richard S. Sutton for their foundational work in reinforcement learning algorithms.

• The $1 million prize recognizes their contributions in “introducing the main ideas, constructing the mathematical foundations, and developing important algorithms” for one of AI’s most significant approaches.
• Both scholars developed theories of how computers could formulate internal models of their environments and learn through trial and error—concepts that have powered some of AI’s most impressive achievements.

How reinforcement learning works: The technique resembles how a mouse might navigate a maze, learning which actions lead to rewards (cheese) and which lead to dead ends.

• Programs using this approach initially act somewhat randomly, trying different moves in their environment and receiving either positive or negative feedback.
• Based on this feedback, the system develops an “internal model” that helps it estimate potential rewards for various actions, eventually formulating a “policy” to guide future behavior.
• Success requires balancing exploration of new choices with exploitation of known good options—neither strategy alone is sufficient.

Major applications: Reinforcement learning techniques have powered some of AI’s most impressive achievements in game mastery.

• Google DeepMind’s AlphaZero used this approach to achieve mastery in chess, shogi, and Go in 2018.
• Similar techniques helped AlphaStar reach “grandmaster” level play in the complex video game Starcraft II.

Important distinction: This classical reinforcement learning differs significantly from the “reinforcement learning from human feedback” (RLHF) used by OpenAI and other large language model developers.

• While both share similar terminology, RLHF is specifically used to shape LLM outputs to be helpful and inoffensive, representing a different technical approach.

Why this matters: Reinforcement learning represents what Sutton has called “the first computational theory of intelligence” at a time when he believes AI lacks agreed-upon theoretical foundations.

• Sutton, who worked as a Distinguished Research Scientist at DeepMind from 2017 to 2023, has emphasized that reinforcement learning provides a crucial theoretical framework for understanding intelligence.
• The research also highlights the importance of curiosity and play in learning—suggesting that setting goals that seem unnecessary at first may ultimately prove valuable for both human and artificial intelligence.