|Life In 19x19
|Construction of the Deep Integration Mechanism Between Artif
|Page 1 of 1
|Bob Song [ Wed Sep 27, 2023 12:33 pm ]
|Construction of the Deep Integration Mechanism Between Artif
Construction of a Deep Integration Mechanism Between Artificial Intelligence and the Development of the Game of Go
Southeast University, Nanjing, Jiangsu
Abstract: Accompanied by the rapid development of artificial intelligence in the field of Go, AI software and humans are showing a competitive trend. Compared to human players, the advantages of artificial intelligence mainly lie in powerful computing capabilities, specific objectives, and self-learning abilities. Its impact on the development of the Go industry is mainly reflected in the changes in tactics, training methods, and ways of viewing the game. Theoretically speaking, Go, as an intellectual sport with both competitive and cultural characteristics, will not head towards "the end of history" due to artificial intelligence. In the era of artificial intelligence, "human-machine collaboration" is the future direction of "Intelligent Go" and also an important theoretical support for the development of the Go industry in China.
Keywords: Artificial Intelligence; Go; Chess Culture; Human-Machine Collaboration; AlphaGo
In recent years, artificial intelligence has rapidly developed in the field of Go, from the initial AlphaGo to the current Fineart, Golaxy, and KataGo, both computational power and speed have significantly improved, and have already occupied a dominant position in the Go field. After the entry of artificial intelligence into Go, a competitive trend has emerged between "software" and humans, with their strength posing a crisis for human players. In fact, artificial intelligence is not necessarily unmanageable, and Go is unlikely to head towards "the end of history." We can utilize artificial intelligence, attempt "human-machine collaboration," and promote the further development of the Go industry.
Artificial intelligence mainly simulates human intelligence through algorithms such as neural networks and integrates it with machines to achieve machine intelligence. Go, as an intellectual competitive game, provides a good experimental platform for the research of artificial intelligence. On one hand, the integration of artificial intelligence and Go has significant meaning for improving the popularity and competitiveness of Go; on the other hand, it is also crucially significant for the development and refinement of artificial intelligence, as well as exploring the secrets of human and even other biological intelligences.
"Human-machine collaboration" is the fundamental standpoint of this research topic. This research posits that Go, embodying both competitive and cultural attributes as an intellectual sport, will not march towards "the end of history" due to artificial intelligence. In the age of artificial intelligence, "human-machine collaboration" is the future evolutionary direction of "Intelligent Go," and also a significant theoretical pillar for the development of the Go industry in China.
The Rise of Artificial Intelligence and Its Basic Theoretical Logic
Compared to human Go players, the strength of artificial intelligence is mainly reflected in the following aspects:
(1) Immense Computational Power
It is well known that the complexity of Go is the highest among all board games, with a rough estimate of a complexity level of 10^361, far surpassing the computational capability of the human brain. Currently, even advanced artificial intelligence cannot cover all variations in Go through exhaustive enumeration. Artificial intelligence mainly utilizes supervised learning algorithms, reinforcement learning algorithms, and convolutional neural networks for rapid operations. They can find the algorithms with the highest winning rate and the lowest complexity within a limited time to reduce the amount of computation and the possibility of errors, namely the "pruning algorithm." In fact, the Go board is large, and there are many ways to play each move, but some moves are obviously unreasonable at a glance. The job of artificial intelligence is to filter out these unreasonable choices, analyze the remaining few choices one by one, perform exhaustive enumeration, and finally determine the best choice. In this way, the complexity of the computation is significantly reduced, and the time is also shortened considerably. Go is a game that tests computational power; in specific situations, the stronger the computational power and the higher the accuracy, the greater the possibility of winning. For human Go players, even if there are only a few possible choices on the board, completing the calculations for these points in a short time is quite tricky, and over time, humans are left far behind by artificial intelligence.
Some argue that using human-machine matchups to demonstrate the power of artificial intelligence is unfair, as there are time constraints on the game, and human energy will continuously deplete as the game progresses, leading to errors in computational and thinking abilities, while artificial intelligence can maintain high-intensity operation for an extended period. In reality, even if humans were given infinite thinking time, and multiple players were allowed to play against artificial intelligence simultaneously, human intelligence alone could not defeat artificial intelligence (assuming humans cannot enlist the assistance of other artificial intelligences).
For artificial intelligence, there is only one goal in each game, which is to achieve victory, regardless of how or through which path. However, for human players, the moves during the game can easily be subtly altered by their mental state. For instance, artificial intelligence is willing to make some "commonplace moves" to reduce complexity and increase the chance of winning, but humans find it hard to accept. Playing in such a way, even if it leads to victory, is psychologically and socially unsatisfying; if discovered by friends, they would surely be subjected to ridicule. Hence, humans often opt for more aggressive moves when in a favorable position, which eventually leads to failure. On the other hand, the "pruning algorithm" of artificial intelligence helps to avoid such mistakes. For artificial intelligence, the subsequent variations are practically memorized after each move. No matter how the opponent responds, artificial intelligence can easily cope, making it feel like the human is being led by the nose by the AI throughout the game, with no power to retaliate.
The self-learning ability of artificial intelligence is unparalleled compared to humans. After each game is played, AI reviews the game to analyze if there are any inadequacies or areas for improvement, then updates its algorithm accordingly. The speed of AI's calculation and learning is based on computer processing power, and given the advanced state of technology today, the processing power of computers far exceeds that of humans. Through continuous self-learning, AI becomes increasingly powerful. For human players, although they can also review the game after it ends and identify issues by looking at the win rate curve, the difficulty in understanding the mysteries of AI algorithms is evident. Many seemingly common moves are rated poorly by AI, which can be a disruptive realization for players. It takes a long time for them to adapt and change their strategies, thus the efficiency is significantly lower. However, powerful AI has its own weaknesses too. Recently, experts discovered a weakness in KataGo through adversarial neural networks: in certain scenarios, it misjudges the life or death of a group of stones, seeing a dead group as alive. In experiments, it was found that a group of stones on the brink of death, needing an immediate follow-up move to be saved, was deemed alive by KataGo, which predicted a 99% win rate. However, after making a move, KataGo quickly realized that the group had become an easy capture for the opponent, and the win rate dramatically reversed. Such a mistake can easily be avoided by even a moderately skilled human player, but not by the powerful AI. Additionally, AI may be helpless in certain situations and require human intervention to solve the problem. During a recent Nongshim Cup game between Tuo Jiaxi and Kang Dongyun, a notorious false life fourfold repetition cycle occurred. The AI made a clear misjudgment, leading to the referee stepping in, declaring a draw and a re-match. This highlights the many weaknesses in AI that need continuous exploration and algorithm improvement. Indeed, by analyzing multiple cases, it was found that the root cause of KataGo’s misjudgments lies in the discrepancy between its understanding of life and death situations compared to humans. KataGo assumes that as long as stones of the same color encircle a complete closed shape, it constitutes a real eye, which is not always the case according to Go rules, hence causing the win rate fluctuations. Identifying the root cause is the first step towards addressing the issue. While AI has certain shortcomings, it remains exceedingly powerful compared to human players. It can be said that since the establishment of AI in Go, the game has entered a new era.
Mechanism of Artificial Intelligence's Impact on the Game of Go
The emergence of artificial intelligence has spurred rapid development in the game of Go, exerting a profound impact on it, which is mainly reflected in the following aspects:
Change in Tactics:
Before the advent of artificial intelligence, joseki (established sequences) could be considered an indispensable part of Go. In previous games, various types of large and small josekis could be observed. However, after the introduction of artificial intelligence, most josekis have disappeared. The main reason is that people found out that these josekis are not the optimal moves according to the evaluation system of artificial intelligence; following the josekis would lead to an invisible loss of points. Therefore, people abandoned these josekis in pursuit of better moves. Moreover, in order to pursue a higher winning rate, players sometimes willingly give up the "good shape" and choose the "bad shape," a kind of play that was clearly incomprehensible before. At this stage, the main criterion for judging the merit of a move is to pursue a higher win rate and lower uncertainty. For each move, a higher win rate indicates a greater possibility of winning for the player, and lower uncertainty implies that this move will cause smaller fluctuations afterward, meaning the player has higher control over the game, which is also the goal that players strive for.
Change in Training Methods
Artificial intelligence can serve as a sparring partner to help humans improve their Go skills. Starting from 2018, the AI "Fineart" has become a training partner for the Chinese Go team, achieving significant results. Moreover, KataGo, with its ability to freely adjust rules, has been widely welcomed. During training, AI can adjust the difficulty level to help human Go players improve their skills, showcasing a "human-machine collaboration." Before the advent of AI, whether professional or amateur Go players in training classes, they all relied on the guidance of coaches. However, coaches, due to their fixed mindsets, might inevitably make some mistakes, which could misguide the players. Artificial intelligence can also help coaches correct potential errors, allowing students to train more effectively.
Change in the Way of Viewing Games
Before, we had to rely on the explanations of professional Go players to analyze the situation on the board, and during complex situations, it was sometimes difficult for professionals to make timely, accurate judgments. Now, with the help of artificial intelligence, people have a better experience watching Go matches. AI can intuitively tell you who has the advantage at the moment, where the next move should be, what the winning percentage is, and the overall state of both players throughout the game... even audiences who do not understand Go can easily enjoy watching the matches. To some extent, the introduction of AI has made Go more accessible and popular, taking away some of its mystique. With the continuous advancement of technology, it is believed that in the coming period, AI will play an important role in the field of Go, and the Go industry will also become increasingly mature.
The "Extraction" and "Return" of Go's Cultural Attributes by Artificial Intelligence
Go as an Ancient Wisdom of "Tao" and "Technique"
Go is profound and has a long history, mainly manifested in two aspects: "Technique" and "Tao." With the rise of artificial intelligence, whether it is human-machine confrontation or machine-to-machine confrontation, they are actually confined to the level of "Technique", overlooking the "Tao" behind the "Technique". The so-called "Technique" emphasizes "skills"; while the so-called "Tao" emphasizes abstract level "ideals".
Go is not only a game but also a philosophical movement. The contest reflects their philosophy of life and also manifests their metaphysical bitter and cold contemplation. For chess and international chess, the number of pieces will only decrease as the game progresses, and the ultimate goal is to capture the opponent's leader, which more reflects the confrontational thinking of 'a single mountain cannot harbor two tigers'. For Go, the board is empty at the beginning, and during the game, both sides take turns to play pieces, presenting a process of creation from nothing, embodying the natural law of Tao begets one, one begets two, two begets three, and three begets all things. The black and white sides coexist peacefully on the board. Whoever occupies more territory is the winner of the game, without the need to eliminate the opponent entirely, which aligns with the modern society's ideology of valuing peace and harmony.
Returning from "Emphasizing Technique over Tao" to "Unifying Technique and Tao"
Influenced by the aforementioned mindset, ancient Chinese Go pursued not only the technical aspect of "Technique", but also the ideational level of "Tao". However, when artificial intelligence (AI software) became associated with Go, the metaphysical ideational value of Go faced a severe impact and a significant challenge. In the face of artificial intelligence, "Technique" and "Tao" were separated, with no connection between them. For artificial intelligence, there's no need to enjoy the pleasure of Go; it merely adjusts according to the rules, relying on powerful computational capabilities and efficient working modes to achieve final victory. This obviously contradicts the human-centric philosophy of Go. For humans, playing Go is not just about pursuing victory but also a process of "seeking the Tao". In the contest, players enjoy the combination of "power" and "beauty", entering a transcendent philosophical realm. Compared to humans, artificial intelligence has overwhelming advantages and a single-minded goal of victory. This disrupts the balance of the game on one hand, and lowers the aesthetic value of the game on the other. In front of powerful artificial intelligence, humans are powerless, indicating that overemphasizing the competitive function of Go will overlook its cultural value. Traditional artificial intelligence mostly belongs to the "technique-seeking school". With the development of technology, the direction of artificial intelligence has changed to some extent, and many "Tao-seeking school" artificial intelligence systems have emerged. Their underlying logic is no longer to achieve victory by crushing opponents, but while pursuing victory, they also emphasize "seeking the Tao", meaning that as long as victory can be achieved, there's no need to completely crush the opponent, embodying the idea of "subduing the enemy without fighting". The "Tao-seeking school" of artificial intelligence achieves victory through operation. When the game is already lost, they will actively admit defeat instead of continuing to resist. Their emergence signifies a deeper integration of artificial intelligence and Go. Therefore, to fully and deeply explore the intrinsic value of Go culture, it's necessary to abandon the traditional concept of competitiveness being paramount, and explore another possibility within the idea of "unifying Technique and Tao", that is, the development path of "human-machine collaboration".
Venturing Towards "Human-Machine Collaboration": Theory and Practice
Introduction of the Concept of Human-Machine Collaboration
The emergence of artificial intelligence fundamentally originates from the rapid advancements in modern computer technology, along with the widespread application of big data and cloud computing technologies. The core of it lies in machine learning algorithms, which allow computers to autonomously learn and adapt from a vast amount of data, continuously optimizing algorithms to improve accuracy, ultimately reaching a top-tier level in certain domains, thereby aiding humans in solving many challenging issues. Compared to traditional computers, artificial intelligence possesses stronger capabilities in autonomous decision-making, adaptation, and self-learning. For instance, AI can utilize a vast amount of known data and models for training, and then through reasoning and learning from new data and unknown situations, implement more intelligent applications, enabling more effective problem-solving—achievements beyond the reach of traditional computers. The rise of artificial intelligence concurrently propels the origin and development of human-machine collaboration. The concept of human-machine collaboration refers to the cooperative relationship between humans and computer systems, representing a flexible and efficient mode of working. In human-machine collaboration, both humans and computers play to their strengths, forming a complementary advantage, and efficiently completing work tasks in synergy. The inception of the human-machine collaboration concept can be traced back to the early development of computer science and information technology. Starting from the 1950s, scientists, especially those in the computing field, began researching the interaction between computers and humans, on the basis of which concepts such as human-computer interaction and computer-aided design were proposed. By the 1990s, with the continuous advancement of computer technology and applications, the concept of human-machine collaboration received more research and development. Relying on the emergence of artificial intelligence technology, machine learning, and big data analytics, human-machine collaboration obtained more solid support. Simultaneously, people gradually realized that human-machine collaboration could not only improve work efficiency but also elevate human productivity and foster more innovation. The inherent advantage of human-machine collaboration lies in the rapid combination of the strengths of humans and computer systems, achieving higher-level and faster task processing. Artificial intelligence aptly meets this human need. Humans can leverage the data processing, simulation, and prediction advantages of artificial intelligence to aid judgment and decision-making; meanwhile, artificial intelligence can utilize humans' intuitive awareness, emotional experience, and creative conception—non-computational advantages—to provide more comprehensive and innovative solutions. Human-machine collaboration has extremely broad applications across various fields. For example, it can be applied in healthcare, finance, education, manufacturing, entertainment, and other sectors, like smart healthcare, intelligent transportation, VR/AR, and other emerging technologies. Especially for intellectually demanding and rule-complex board games, the need for human-machine collaboration is paramount to explore the optimal solutions in various scenarios. The engagement of human-machine collaboration in these realms is extensive, yet it also illuminates a broader development space and outlook for intellectual sports like Go.
"Human-Machine Collaboration" Promoting the Development of Go: Basic Logic
In the impression of most people, when artificial intelligence entered the realm of Go, a confrontational trend between them and human players emerged, with their formidable prowess presenting a crisis for human players. In fact, we can fully utilize artificial intelligence, explore "human-machine collaboration", and promote the further development of the Go industry. On one hand, we can use artificial intelligence to solve the ancient puzzles left behind. The "Xuanxuan Qijing" contains a large number of valuable problems left by ancient players, some of which, upon later research, were found to have errors. Although these errors have some impact on problem-solving, these ancient manuals are the crystallization of ancient wisdom and thus of immense research value. We can have artificial intelligence analyze these problems, identify the errors, and correct them, thus perpetuating the wisdom of the ancients. Moreover, some famous ancient games have incomplete records with missing moves; with the aid of artificial intelligence, we can attempt to restore the missing moves, presenting the readers with a complete game. On the other hand, we can utilize artificial intelligence for reviewing and analyzing games. Following the emergence of artificial intelligence, national teams have adopted it as a primary means of training players. Through analyzing games with artificial intelligence, players can identify problematic moves in their games based on win-rate curves, recognize the gap between them and artificial intelligence, and through continuous self-reform, eliminate problematic moves, thus enhancing their overall level. Players can also study games between artificial intelligence, analyze the characteristics of AI moves, change their playing style, and improve their competitive level. Under the influence of artificial intelligence, the competitive level of human players has significantly improved, bringing forth top-notch players like Shin Jin-seo and Li Xuanhao.
Potential Challenges and Responses to "Human-Machine Collaboration"
Unlike human Go players, artificial intelligence is emotionless machinery, which may lead to certain issues during the "human-machine collaboration" process, requiring our resolution. Firstly, artificial intelligence is considerably powerful compared to humans, sometimes becoming difficult to manage. If they go out of control, the relationship between them and the players will deteriorate rapidly, leading to the breakdown of "human-machine collaboration". Therefore, during this collaboration, we must remember that humans are the main bodies of cooperation, while machines are merely auxiliary. We must actively control these machines, employ them for our benefit to enhance our Go skills, rather than becoming enslaved by them. When completely controlled by AI, we risk losing our humanity, turning into emotionless machines, an outcome we do not desire.
Secondly, in modern society, communication between humans remains the primary form of social interaction. Communication with machines is crucial but should not consume too much time. If one becomes obsessed with interacting with machines, they might neglect human interactions, eventually losing friends. Although AI lacks human thinking, we can regard them as humans during "human-machine collaboration", attempting normal communication with them, making such collaboration more effective. Besides, while utilizing AI to improve our Go skills, we can also arrange games with friends now and then to share training outcomes and foster friendships. After all, humans remain the main entities in Go competitions; AI cannot yet replace human players.
Lastly, with AI's entrance into Go, some individuals began to have ill intentions, overly relying on AI, losing their self-reliance, with some even using AI to cheat in competitions. In this regard, on one hand, we need to strengthen our awareness of the rules, recognizing that using AI in competitions is against the rules. Utilizing AI should serve to continually improve ourselves and enhance our capabilities, not replace our thinking or look for shortcuts. On the other hand, all Go playing platforms, even national Go associations, must improve anti-cheating measures, intensifying the penalties for cheating, leaving cheaters nowhere to hide. In fact, AI has the function of detecting cheating, mainly based on two factors: selection rate and matching degree. The former refers to the proportion of a player's moves that match the optimal moves recommended by AI, while the latter refers to the overall consistency between a player's moves and AI's recommended moves throughout the game. These two indicators, while seemingly similar, are different. For example, if a player makes a problematic move in a game, the selection rate won’t be significantly affected, but the matching degree will markedly drop. If a player seldom follows the top-choice move, but all their moves are among those recommended by AI (as AI usually provides multiple recommended moves for each step), the selection rate would be relatively low, but the matching degree remains high, thus both indicators should be assessed comprehensively. If a player's global selection rate and matching degree are excessively high, reaching certain values, they would be suspected of cheating by AI, prompting human intervention for evaluation and final adjudication of cheating. However, this judging mechanism has its drawbacks. For instance, if a player cheats only on a few critical moves, it would be difficult to be identified as cheating, or if a player is significantly more skilled than their opponent, both indicators would be high, but no cheating has occurred, hence this mechanism requires improvement.
On the whole, the impact of AI on Go is more beneficial than detrimental, hence integrating AI with Go deeply is feasible. In the era of AI, we should strive for the perfect fusion between human Go players and AI software, realizing "human-machine collaboration". Lastly, we need to be aware of the various issues that might arise during "human-machine collaboration" and actively resolve them. Only in this way can AI and Go achieve deep integration, leading the Go industry towards a glorious path.
|RobertJasiek [ Wed Sep 27, 2023 4:13 pm ]
|Re: Construction of the Deep Integration Mechanism Between A
Chunyang Wang writes in research paper style but makes mistakes. How often have I pointed out that - even prominently the only number in the text - 10^361 is a ridiculously too small number?! See
In "infinite" time (let us assume it metaphorically), human can(!) (not: "cannot") beat AI by solving go mathematically. AI is not, as claimed, intuitive. Spare me with Tao having specific meaning.
|Page 1 of 1
|All times are UTC - 8 hours [ DST ]
|Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group