Did AlphaGo Zero Really Breakthrough in Algorithm Design?

AlphaGo made a significant impact in the world of artificial intelligence, particularly in the field of Go. However, when we compare AlphaGo with AlphaGo Zero, the latter's approach might not be seen as a major innovation in algorithm design.

AlphaGo vs. AlphaGo Zero: A Tale of Paradigm Shifts

AlphaGo, the groundbreaking AI, was trained on a large dataset of human games and plays. This approach gives the AI a thorough understanding of the game but also carries the inherited biases of human play. On the other hand, AlphaGo Zero started from a tabula rasa, teaching itself solely through reinforcement learning without any pre-existing knowledge of the game. While this process was slower, the results were surprisingly profound.

The primary difference lies in the starting point. While AlphaGo benefited from the vast repository of human wisdom, AlphaGo Zero developed its knowledge purely through experience. This approach was more akin to the way a human would learn the game, without any prior knowledge.

Novelty and Enlightenment in AI

I often reflect on the AI koans from the MIT AI Lab, which remind us of deeper truths in the spiritual and intellectual journey of AI. For instance, the story of “Sussman Attains Enlightenment” is a fascinating reminder of the limits of pure randomness in AI.

Sussman attains enlightenment: Sussman, a novice at the MIT AI Lab, was training a randomly wired neural network to play a simple game called Tic-Tac-Toe. Minsky, the mentor, asked why the network was randomly wired. Sussman replied that he wanted the network to have no preconceptions. However, the key moment of enlightenment came when Minsky suggested that Sussman close his eyes, implying that the room itself would be empty, signifying a complete starting point.

Koans are not meant to be explained but to provoke thought and reflection. In the context of AlphaGo Zero, the tabula rasa approach can be seen as akin to Minsky’s suggestion for Sussman. However, as the story goes, Sussman found enlightenment when he realized the crucial aspect of a clean and empty slate.

Lessons from AlphaGo and AlphaGo Zero

The larger question that arises from these examples is what we can learn about using AI to solve other problems. Both AlphaGo and AlphaGo Zero highlight the importance of different approaches in achieving goal-directed intelligence. AlphaGo benefited from a rich training dataset, while AlphaGo Zero relied entirely on self-play and reinforcement learning.

This duality in approach could guide future AI development in other fields. Just as Sussman found enlightenment through his own exploration, we can see the value in both leveraging existing knowledge and starting afresh to solve new problems.

Ultimately, the real breakthrough is not in the method itself but in the principle that there are multiple paths to achieving intelligent solutions. Both AlphaGo and AlphaGo Zero have shown us the immense potential of AI, and the lesson they offer is that it is the journey through which we can unlock these potentials.