AlphaGo Zero Compared to Its Predecessors and Performance

Posted: August 27th, 2021

Student’s Name

Instructor’s Name

Course

Date

AlphaGo Zero Compared to Its Predecessors and Performance

AlphaGo Zero is a new program. Compared to its predecessors, AlphaGo Zero is considered the most powerful program. It is advanced, with the capability to perform tests. It surpasses Lee’s version, characterized by generating new own ideas for over 2000 years old games (“AlphaGo Zero: Starting from Scratch”). Equally, the version illustrates new tricks for a teaching machine, hence less reliable on a human being. For example, during testing, the software could have consumed over 160,000 amateur game data sourced on the Go community. Afterward, the developers acknowledged the machine playing over one million games against itself (“AlphaGo Zero: Starting from Scratch”). Thus, unlike its predecessors, AlphaGo Zero is less reliant on human interventions for its performance.

The AlphaGo Zero predecessors include AlphaGo Fan, AlphaGo Lee, and AlphaGo Master. AlphaGo Fan has 176 distributed Graphics Processing Units with an Elo rating of 3,144. It matches 5:0 against Fan Hui. Subsequently, AlphaGo Lee is designed with 48 distributed Tensor Processing Units with an Elo rating of 3,739. It matches 4:1 against the Lee Sedol. Lastly, AlphaGo Master is made of a single machine with 4 Tensor Processing Units and an Elo rating of 4,858. It matches 60:0 against the professional players in the Future of Go Summit. However, the evaluation of AlphaGo Zero shows that it is developed as a single machine with 4 Tensor Processing Units and an Elo rating of 5,185 (“AlphaGo Zero: Starting from Scratch”). The version matches one-hundred to zero against the professional players and eighty-nine to eleven against the AlphaGo Master. Thus, these differences gave the version more power than its predecessors. This has enhanced its performance by being self-taught with limited reliance on human beings. The changes also enhanced its capacity to absorb extensive data and perform several games simultaneously, unlike predecessors.

Works Cited

“AlphaGo Zero: Starting from Scratch.” Deepmind, deepmind.com/blog/article/AlphaGo-zero-starting-scratch.