The Battle of Chess Engines: AlphaZero vs. Stockfish

The world of chess has been witnessing a growing debate over the supremacy of various chess engines. Among them, AlphaZero and Stockfish have been at the center of attention. AlphaZero, a self-taught chess-playing program, made headlines in December 2017 by defeating the highly acclaimed Stockfish engine in a series of matches. However, recent events and advancements suggest that the narrative is more complex than initially thought.

AlphaZero's Triumph in 2017

In December 2017, DeepMind released a groundbreaking paper detailing how AlphaZero, a neural network-driven chess engine, achieved remarkable results. Developed without human guidance, AlphaZero managed to win 28 out of 100 games against Stockfish, without losing a single game and drawing the rest. This victory highlighted the immense potential of reinforcement learning and neural networks in complex strategic games.

Training and Performance

DeepMind estimated that AlphaZero reached a higher Elo rating than Stockfish after only four hours of training. After nine hours of training, the algorithm defeated Stockfish in a time-controlled 100-game tournament, securing 28 wins, 72 draws, and 0 losses. These results were a testament to AlphaZero's advanced playing style, which emphasized long-term strategic planning and creative tactics, contrasting with Stockfish's more traditional evaluation-based approach.

Hardware Considerations and Debates

Some debates have centered around the hardware differences between the two engines. AlphaZero was running on a powerful mainframe, whereas Stockfish was likely running on less powerful hardware. This disparity raises questions about the fairness of the match. Additionally, it has been suggested that Stockfish might not have been allowed to use all its features, which have been implemented to make it play more strongly. While no conclusive proof of such allegations has been provided, this remains a point of contention in the chess community.

Computer chess software is highly dependent on the hardware it runs on, further complicating the comparison. If Stockfish were running on similar hardware, the results could have been vastly different. The unequal hardware makes it difficult to determine the true capabilities of each engine.

Leela Chess Zero and Subsequent Developments

In TCEC season 17, the “Leela Chess Zero” implementation defeated Stockfish in the final with a score of 17-12 and 71 draws. TCEC is known for its competitive openings and varied test scenarios, making it an excellent benchmark for engine strength. This win underscores the ability of neural network-based engines to excel in complex environments.

Stockfish developers responded to the Leela Chess Zero victory by enhancing SF with a neural network of its own, creating “Stockfish NNUE”. This update significantly improved Stockfish's performance, and it went on to win seasons 18-20. This shows that the chess engine world is in a constant state of evolution, with each engine striving to outperform the others.

The ongoing development and improvement of chess engines highlight the dynamic nature of this field. While AlphaZero and Stockfish have set benchmarks, the future of chess engine development is likely to be characterized by continuous advancements and innovations.