Google’s AlphaGo Zero a self-learning model thrashes Older AlphaGos

Oct. 21, 2017, 9:48 a.m. By: Vishakha Jha

AlphaGo Zero

Identification of Opportunities leads us to change for a better tomorrow. With the same thought process, Deep Mind has brought up AlphaGo Zero, a revolutionary computer program which has been claimed to be better than the previous version which has already defeated the best of human minds. Deep Mind is a subsidiary of Google, the world's leading industry in artificial intelligence research and its application for positive impact. One of the key features which distinguish it from rest of its category is that it works on the concept of self-learning i.e. It will learn simply by playing alone, and against itself. This new advancement brings us closer to the onset of general purpose algorithm which is expected to have the ability to solve some hardest problems in the field of science making our life simpler and providing us ease of access.

The original AlphaGo is an AI developed by Deep Mind which is known for its historic victory and becoming the first ever computer program to beat a top-level Go player. The game of Go was originated in China over 3000 years back. It has still maintained its existence due to its profound complexity and the intuitive nature. Due to its level of complexity, it has been a challenge for the world of AI. AlphaGo works by the combination of two very crucial concepts- advanced tree search with deep neural networks. The understanding of the game is developed by analyzing a large number of games and learning from its mistakes through a process known as reinforcement learning.

As of all mankind appreciates faster and accurate work. The AlphaGo Zero marked its existence through a paper which was published in the journal Nature and is claimed to be better than the best. The AlphaGo Zero is said to be a great step towards advancement because it has mastered the game from scratch within a short span of time without any human interference except providing basic rules of the game as input. AlphaGo Zero won 100 to 0 against Lee Sedol, the South Korean legend. The reason behind highlighting its success is due to its autodidact nature which could further lead us towards the improvement in a number of dimensions. The triumph marks a milestone on the way to general AI algorithm which has the ability beyond the boundaries of Go game table.

At DeepMind, AlphaGo Zero is already being experimented on the aspect of how proteins fold and expected to play a major role in various other experimental pursuits. The basic difference between the previous version and Zero version includes-

  • The difference in approach of game initiation where AlphaGo zero requires only game's basic rule as input whereas previous versions dependent a small number of hand-engineered features.

  • AlphaGo Zero, unlike the rest, does not rely on rollouts but it works on neural networks to evaluate positions efficiently.

  • Unlike the previous version, the AlphaGo Zero does not require two different approaches for selecting the next move(Policy network) and another to calculate the winner(Value network) rather it combines both the moves efficiently.

It is claimed to be more powerful than previous approaches because it is independent of human expertise and has the ability to work on knowledge itself. After each game, it revises its neural network, making it stronger player for the next turn.

"I hope that these kinds of algorithms and future versions of AlphaGo-inspired things will be routinely working with us as scientific experts and medical experts on advancing the frontier of science and medicine," Hassabis, CEO of Deep Mind.

These moments of success not only boost our confidence but make us realize the potential that technology possesses to positively influence the society.

AlphaGo Zero: Starting from scratch

Video Source: DeepMind

Image Source: 9to5Google