How AlphaGo defeated a go master

When South Korean go champion Lee Sedol lost his first game to AlphaGo, the whole world was shocked. It was more than just a game of go: It was a milestone in the development of artificial intelligence (AI).

Exactly how did AlphaGo do it? Why were the go programs that came before unable to do it? And how human-like is this AI? Cognitive psychology might offer some interesting insights on the issue.

Since the 1940s, psychologists have been studying people who excel in certain areas, hoping to understand their psychological structure and explain their development. The first research of this kind focused on chess experts, where cognitive psychologists compared chess experts and chess rookies, and were surprised to find they are not very different. Experts are just faster at finding the right move, and the reason has nothing to do with their intelligence or memory capacity, but their advanced ability to recognize patterns on the chess board.

When looking at the pieces in a game, those who excel at chess memorize the pattern they present, which includes the correlations between various pieces. These correlations can reveal the intentions and strategies of a chess player. As a chess player gains more experience, they also become better at reading and memorizing the patterns, so that when a familiar pattern shows up, an expert can immediately predict the ensuing development of the game based on experience. Rookies, on the other hand, cannot see the relations between the pieces in a game. To them, the way the pieces are placed on the board is random. As a result, they often misjudge a situation.

AlphaGo is different from AIs of the past in a number of ways. First, it uses neural networks to recognize patterns. Neural networks are superior to traditional AI, because they can accumulate and learn from their own experiences. In addition to remembering shapes, they can extract patterns from those shapes, and they can also use Monte Carlo methods — computational algorithms used to obtain results from random sampling — to assist their strategic choices. This means that AlphaGo can make decisions based on shapes and patterns to analyze the situation and make an assessment of which move would increase the possibility of victory.

These mechanisms are characteristics of human cognitive systems, but in addition to relying on these two mechanisms, AlphaGo also possesses computing power that far exceeds the human brain.

We have limited cognitive sources at our disposal. For example, “1+1=?” is a simple problem, but many people would find “12345+56789=?” to be a bit of a challenge. Working memory span — the longest list of items a person can remember correctly immediately after having been presented with the list — is an indicator of cognitive resources, and just like a computer’s random-access memory, this is the brain’s platform for processing information. It is limited in size, and the more complex a task, the more cognitive resources it uses up, and this has a direct effect on how fast and how accurate human calculations are. Compared with AlphaGo, Lee’s greatest weakness was that he is human, and thus is restricted by his working memory span.

Although Lee is also capable of analyzing a situation and then arriving at an estimate, his working memory span is limited, and that affects the number of moves available for him to choose from. From this perspective, AlphaGo is indeed different from humans, but in the field of AI, similarity to humans is not necessarily the main point. Instead, the goal is to create a robot capable of displaying intelligence.

From this perspective, AlphaGo clearly set a new milestone in the field of AI, although the very fact that its intelligence surpasses the human brain shows that it is not sufficiently human. For humans to be able to beat it, perhaps it must be made more human.

http://www.taipeitimes.com/News/editorials/archives/2016/03/20/2003641982

who-AlphaGo and  South Korean go champion Lee Sedol

what-AlphaGo can make decisions based on shapes and patterns to analyze the situation and evaluate which moves increase the likelihood of success.

how- South Korean go champion Lee Sedol lost his first game to AlphaGo

Keywords-excel limited humams complex computing