AI Scholars Awarded Turing Prize for Technique Behind AlphaGo's Chess Victory

In the past decade, artificial intelligence has dazzled us with its advancements, particularly through a technique where computers make random choices and learn from the outcomes. This method, known as reinforcement learning, has been pivotal in achieving remarkable feats in AI.

Take Google DeepMind's AlphaZero program from 2016, which by 2018 had mastered the complex games of chess, shogi, and Go. Similarly, AlphaStar used this approach to reach "grandmaster" level in the video game *Starcraft II*. These achievements highlight the power of reinforcement learning.

On Wednesday, the field celebrated a significant milestone as two AI scholars were honored for their groundbreaking work in advancing reinforcement learning. Andrew G. Barto, a professor emeritus at the University of Massachusetts, Amherst, and Richard S. Sutton, a professor at the University of Alberta, Canada, received the prestigious 2025 Turing Award from the Association for Computing Machinery (ACM).

Recognition of Pioneers in Reinforcement Learning

The ACM praised Barto and Sutton for laying the groundwork for reinforcement learning, stating that they "introduced the main ideas, constructed the mathematical foundations, and developed important algorithms." This accolade, which comes with a $1 million prize, is often seen as the computer industry's equivalent of a Nobel Prize.

Reinforcement learning can be likened to a mouse navigating a maze to find cheese. The mouse learns which paths lead to progress and which to dead ends. Similarly, neuroscientists believe intelligent beings, like mice, develop an "internal model of the world" to guide their actions.

Sutton and Barto proposed that computers could also develop such internal models. In reinforcement learning, the computer gathers data about its environment—be it a maze or a chessboard—and initially acts randomly. It receives feedback in the form of rewards or penalties, which helps it estimate the outcomes of different actions. Based on these estimates, the program develops a "policy" to guide future decisions, balancing the exploration of new actions with the exploitation of known successful ones.

The Role of Exploration and Exploitation

At its core, reinforcement learning requires a delicate balance between exploring new possibilities and exploiting known strategies. Neither approach alone is sufficient for success.

For those interested in diving deeper, Sutton and Barto's 2018 textbook on the subject is a valuable resource.

It's worth noting that the term "reinforcement learning" is sometimes used differently by companies like OpenAI, which employ "reinforcement learning from human feedback" (RLHF) to refine the outputs of large language models like GPT. However, this is distinct from the method developed by Sutton and Barto.

Reinforcement Learning as a Theory of Thought

Sutton, who was a Distinguished Research Scientist at DeepMind from 2017 to 2023, has argued that reinforcement learning is not just a technique but a "theory of thought." He has expressed concern about the lack of a computational theory in AI, asserting that "reinforcement learning is the first computational theory of intelligence."

Beyond its technical applications, reinforcement learning may also shed light on creativity and free play as expressions of intelligence. Sutton and Barto have highlighted the role of play in learning, suggesting that curiosity drives exploration. Sutton has emphasized that play can involve setting goals that may not be immediately useful but could prove beneficial later.

"Play is a big thing," Sutton remarked, indicating its significant role in the broader context of learning and intelligence.

The journey of reinforcement learning, from its foundational work by Barto and Sutton to its application in games and beyond, continues to push the boundaries of what AI can achieve.