In this blog post, we will look at how far artificial intelligence can approach human intuition through the match between AlphaGo and Lee Sedol.

 

Lee Sedol vs. AlphaGo, the match between humans and artificial intelligence

One of the biggest issues in 2016 was the match between Lee Sedol 9th Dan and Google DeepMind’s AlphaGo. Lee Sedol 9th Dan was known as the strongest player in the world of Go, and Google DeepMind was a subsidiary of Google, one of the world’s most innovative IT companies. The confrontation between the two beings went beyond a simple game of Go and formed a traditional yet interesting confrontation between “computer vs. human,” which attracted a lot of attention. This confrontation was seen as an opportunity to test the limits of existing artificial intelligence and raised questions about the limits of the technology and human capabilities of both sides.
The match was played in five games, and as a result, AlphaGo defeated Lee Sedol 9th Dan with an overwhelming score of 4:1. After the match, people thought that it was practically impossible for humans to beat a computer that quickly calculates all possible outcomes. Even though Lee Se-dol 9th Dan won only one game, many people praised him as a monumental challenger. If AlphaGo had simply relied on calculating all the possible outcomes, this victory would have been just the result of a computer with improved computing power. However, AlphaGo’s victory did not come from an improvement in the performance of its hardware, but from an innovative advancement in its internal algorithms, which showed that AlphaGo’s understanding of Go was on a different level from that of existing AI. AlphaGo’s algorithm has since been applied to various fields, opening up the possibility of having a profound impact on our lives.

 

Existing algorithms: Minimax algorithm and limitations

To understand why AlphaGo is special, it is necessary to first understand the minimax algorithm used by existing board game AI. AI has long been able to defeat humans in games such as Go and chess, and in chess, it has been decades since a computer beat the world champion. The minimax algorithm used at the time is based on the concept of “considering all possible outcomes and choosing the best one.” As the name suggests, the Minimax algorithm chooses the best outcome for itself in preparation for the worst possible outcome for the opponent. In the case of a chessboard, the number of possible moves is relatively small when considering the movements of specific pieces within the limited range of 8×8, so as long as the computing power is sufficient, the player can make moves with a view as far as possible.
However, the number of possible moves in Go is on a different level. The Go board has a large range of 19×19, and there are almost no restrictions on where the stones can be placed, so the number of possible moves from the first move is in the hundreds of millions or even trillions. If all the possible outcomes are to be calculated by the Minimax algorithm, it would take 700,000 years to complete all the calculations, assuming that only the first six outcomes are considered, and that one outcome is calculated per second. Even chess was too difficult to analyze all the possible cases using simple calculations, so various shortcuts were used, which was no longer an effective approach for complex games like Go.

 

AlphaGo’s innovation: intuitive decision making through deep learning

The reason AlphaGo was able to defeat humanity in Go was because it introduced a new approach that overcomes the Minimax algorithm. In Go, it is very difficult to choose the best move, so AlphaGo used a method that uses a decision-making process similar to human intuition to reduce the computational process to the extreme. The core of this new approach comes from a technique called Deep Learning. Deep Learning is an AI methodology that mimics the human brain’s neural network, allowing it to solve complex problems through learning and intuition.
Deep learning is a learning process that involves processing input data through multiple layers of neurons using artificial neural networks. In this process, AlphaGo did not simply calculate all the possible outcomes, but instead used the learned patterns to predict the number of winning moves with a high probability at a particular position and made a move accordingly. Through gradient descent, one of the important learning techniques of deep learning, AI can make more accurate judgments by reducing errors through repeated learning. This deep learning technique has the advantage of being able to perform all operations in matrix form, enabling parallel processing, which allows operations to be processed at an incredible speed using a GPU.

 

Introduction of CNN: Recognizing the Go board as an image

Among the deep learning techniques used by AlphaGo, the most important was the convolutional neural network (CNN). CNN is an algorithm that originally showed great results in image recognition and classification tasks, and is excellent at recognizing patterns in images and extracting features. CNN works by analyzing each pixel of an image and classifying the shape and characteristics of the objects contained within it. AlphaGo learned by imaging the arrangement and layout of Go stones by replacing the Go board with the pixel data of a CNN. Based on the characteristics of a CNN, AlphaGo was able to analyze each move on the Go board like a color pattern in an image and calculate the probability of winning the next move. This allows for highly intuitive numbers to be placed through learning and pattern recognition rather than simple arithmetic, and it allows for decisions to be made on a different level than existing AI.
AlphaGo was equipped with a self-learning function that differentiated it from existing AI, and it developed a judgment ability close to human intuition by repeating training through its own records. This means that deep learning technology has given AI the ability to self-learn, enabling it to perform tasks that require complex thought processes beyond simply solving given problems. This self-learning function is an example of how AI is increasingly mimicking the way humans think and approaching the stage where it can create new strategies on its own.

 

The future that deep learning will open: AI entering our lives

The impact of deep learning techniques on our daily lives is beyond imagination. Even now, artificial intelligence is helping us in many ways by mimicking human thinking and judgment in various fields, such as photo classification, language translation, and speech recognition. If deep learning becomes more widespread in the future and AI can learn a person’s behavior patterns, AI will not be a simple assistant, but will predict a person’s behavior, reduce mistakes, and make the necessary decisions on its own. For example, if we enter an era in which AI learns about people’s health and daily habits to detect risk factors in advance, or predicts traffic conditions in real time to guide the best route, our lives will become more convenient and safer.
Furthermore, as AI becomes more deeply involved in our lives and takes on an increasingly larger role, we will be able to experience new forms of life that we could never have imagined in the past. In this respect, the 2016 match between AlphaGo and Lee Sedol was more than just a game of Go, it was the starting point for the blurring of the line between humans and AI, and it can be seen as a preview of our future life with AI.