游戏里的智能革命：AI如何与游戏共创未来？

The author Wang Shu, a postdoctoral fellow at Tencent Research Institute, pointed out that there is great potential in the field of game artificial intelligence to realize or create humanoid artificial intelligence. As the starting point of artificial intelligence research, the complexity and diversity of its task scenes provide a guarantee for artificial intelligence to approach human intelligence in breadth, depth and flexibility. At present, with the rapid development of generative and decision-making technology, the development trend of game and artificial intelligence resonance symbiosis is more obvious in the global game summit. In the game industry in the generative field, the respondents are using tools to make game content, and in the decision-making field, the team will launch a general game agent again, which can perform more than a variety of tasks in various game worlds according to the instructions of human natural language. The general practice game based on the game environment provides clear measurement standards for decision-making, and the ability to evaluate decision-making with clear and quantifiable rules in the game can be solved. Solve the problem of lack of artificial intelligence research scenes, greatly improve the efficiency of technical iteration and testing. At present, most decision-making research teams, including others, choose games as training scenes, and are committed to building general-purpose agents in different types of game scenes, and on this basis, the team released an agent named "General Artificial Intelligence", which can understand a wide range of game worlds and perform more than a variety of tasks in various game worlds like human beings. However, the ability of language understanding and transfer learning makes many researchers regard the emergence of the agent as the moment. The basic principle and technical path explained in detail in the technical report define it as a universal game agent team that can be extended and guided in multiple virtual worlds. The team selected a popular online game and an engine-based scene as the training environment of the agent, and collected a large number of human players' behavior and operation data from the game to train the agent in detail. In the process of training, the intelligent person will constantly observe the game image information in the learning screen and combine it with the various operation instructions of the player in the game, and then realize the control of the characters in the game to execute various operation diagrams through keyboard and mouse output. The project overview of the intelligent person is an important milestone for the team in the research field of general artificial intelligence, from Go artificial intelligence and StarCraft-based game to the team based on the big language model. It seems that the decision-making and action ability trained by agents in the game environment is expected to migrate to real-world scenes, providing new ideas and new practices for incubating general artificial intelligence. Long before the release, there have been many research projects on general game agents in the industry, two of which are representative ones released by NVIDIA and released by the team in May to play Atari series games and control real robot hands. The arm stacking building blocks use a kind of large language model architecture, and its training materials include images, texts, joint data of manipulator and other multi-modal data sets. In a study in June, Microsoft pointed out that this kind of large model with multi-modal information is very likely to produce an initial intelligent map. Similar to this, NVIDIA, California Institute of Technology and Stanford and other research institutions have jointly built an agent based on my world game to make players' video encyclopedia and user community information of my world game. To train a general agent who can complete various tasks according to the text prompts in my world game for training materials, he can not only complete some simple programmed tasks, but also complete a series of creative tasks according to simple descriptions, such as building a library and drawing capacity model according to descriptions. These two tasks correspond to two different ideas in artificial intelligence research to solve enough tasks or solve a complex task, but there is also one in the above two studies. Limited, for example, specialized intelligence for a specific game can only complete all kinds of tasks in a single game, but it does not have the ability of transfer learning. Although it has certain transfer learning ability, its main application environment is some games rather than game environment, which is far from the real world scene. At present, it has become the consensus in the industry to train universality based on game environment. In the speech, senior scientists in NVIDIA put forward the concept of basic model, and thought that the next frontier of research would be to shape a virtual world. And the training of this model is also inseparable from the game environment. In China, Tencent has also taken the lead in building an open research platform for multi-agent and complex decision-making. Enlightenment relies on the core advantages of Tencent and the glory of the king in the experimental scene of algorithm computing power to provide academic researchers and algorithm developers with a new breakthrough in the ability of the leading application exploration platform in China and realize the effective integration of large language models and training. The appearance of "The Big Language Model" combines with agent training to achieve the breakthrough of agent's decision-making ability and generalization, which can not only better understand all kinds of game environments, but also perform various tasks in all kinds of game worlds according to natural language instructions like human beings, and has a decision-making ability similar to human beings in decision-making efficiency and ability. The founder and Demis Hassabis even bluntly combined the big language model agent training with game environments in an interview. This field has a huge development prospect, and the research investment in this field will continue to increase in the future. Generally speaking, compared with other characteristics and breakthroughs, it is mainly reflected in the following aspects: first, using game environment for training, but paying more attention to the consistency of agent behavior and receiving instructions. In the team's view, games are an important testing ground for artificial intelligence systems, and like the real world, games are also a rich learning environment with responsive real-time settings and changing goals. The drama agent is similar in that it has also observed and learned a large number of human players' behavior data during the training process. The difference is that the purpose of training is not to beat human players or get high scores in the game, but to learn to obey the natural language instructions issued by human beings in various game environments and make behaviors consistent with the instructions in the game environment. Secondly, the large language model is combined with the agent training and the unified and humanized interactive interface language and environment learning are complementary. By learning natural language, the agent's ability to understand general representations and abstract concepts can be improved. 比特币今日价格行情网_okx交易所app_永续合约_比特币怎么买卖交易_虚拟币交易所平台

注册有任何问题请添加 微信：MVIP619 拉你进入群

文章转载注明 网址：https://netpsp.com/?id=54716