Home Tech Three difficulties in game AI: large samples, high costs, and poor flexibility

Three difficulties in game AI: large samples, high costs, and poor flexibility


Text | Competition After conquering Go and “StarCraft”, DeepMind began to target mobile games. In early May, after DeepMind launched AndroidEnv, it began intensive testing of many mobile games. The latest addition to the experience challenge is the Android version of “The Battle of Polytopia”, which is produced and published by Swedish developer Midjiwan. It is a turn-based civilization strategy game. The background of the game’s story is set on a strange and flat “square” planet. Players can lead one of 12 different civilizations, expand the empire, research technology and become the ruler of the entire planet to achieve victory. The game has single player and multiplayer modes, and can support up to 12 players. Currently, Midjiwan is cooperating with DeepMind to integrate “The Battle of Polytopia” into AndroidEnv as a routine task for the latter. The big bang, is a transformative experience coming? “We found this game to be a particularly interesting challenge because it has many features, such as handling long-term planning, imperfect information, diverse UI elements, and uncertainty.” DeepMind commented. Regarding this cooperation, Christian L vstedt, the general manager of Midjiwan, said that DeepMind is a giant in the field of artificial intelligence. We are very proud and excited to have such a strong partner and become a member of the platform. In the Polytopia player community, there are no shortage of experienced players. After integrating artificial intelligence, it will definitely bring different changes to the player’s experience. As an AI platform, AndroidEnv allows AI agents to customize tasks in the game, such as finding the direction of the park, booking a flight, and even getting the highest score. The AI ​​agent mainly makes decisions based on the images displayed on the screen. It can operate through touch screens and gestures like humans. In theory, AI can help developers to quickly generate content, automatically generate plots, and even promote NPC intelligence and weathering. Take Tencent AI Lab as an example. In terms of competitive game AI, Tencent AI Lab launched Fine Art and Consciousness. The former mainly involves Go and Mahjong in chess and card projects, completing the expansion from a complete information game to an incomplete information game. The latter mainly involves the MOBA mobile game “Glory of the King” and sports football, which are respectively explored and tried in heterogeneous distributed multi-agent and more agents + long cooperation. The author learned that the “King of Glory” project team has been trying various new gameplays, but the results are often not satisfactory. Surprisingly, the participation rate of AI man-machine battles is extremely high, exceeding 10% of players. This means that daily active users reach several million. Officially disclosed data shows that the AI ​​Awareness Challenge has attracted more than tens of millions of players to participate. In addition, based on AI’s professional strategy analysis for game understanding and the comprehensive expression of sound text, Tencent also launched a virtual anchor of Glory of the King. In terms of art resources, Tencent AI Lab involves virtual humans and 3D motion generation. It is reported that the internal team is trying to get a more realistic game performance in action generation to adapt it to various actions, terrains and emergencies. Opportunities and challenges coexist Behind the seemingly transformative experience, it actually takes a lot of costs. A senior person with a background in games and AI told me that AI can indeed effectively enhance the gaming experience. The constraints are that the sample size is large and the training cost is high. A big DAU, high-volume competitive product like “Glory of the King”, of course, can be a try. For medium-sized products, it is often not very cost-effective. An executive from a start-up game company said that if the small sample model works well, it is not impossible to try tens of millions. In other words, the large sample model discourages companies. Currently, the industry often discusses AGI (Artificial Intelligence in General), but product flexibility is still a big problem. DeepMind is undoubtedly the most representative when it encounters cost and flexibility dilemmas. It is estimated that the overall training cost of AlphaGo is as high as 35 million U.S. dollars, which consumes enough energy to support 12,760 human brains to work non-stop for three days. In the much-watched “StarCraft” project, DeepMind also encountered similar problems. In the early morning of February 28, 2019, DeepMind’s game AI AlphaStar defeated one of the most powerful professional StarCraft players in the world with a score of 5 to 0. At that time it was regarded as the last ground to break through human intelligence. Under normal circumstances, training AlphaStar requires the use of Google v3 TPU to support thousands of “StarCraft II” to work together. AlphaStar played for a total of 14 days, and each agent used 16 TPUs. During the training, each agent has experienced real-time games for 200 years. Roughly estimated, the training cost reached millions of dollars. Regardless of cost, AlphaStar’s flexibility has also been questioned. Gary F. Marcus, a professor in the Department of Psychology at New York University, believes that the above-mentioned scheme has many limitations. When the same “race” is used in a battle on a single map, its performance is indeed better than that of humans. However, once different “races” are used on other maps, the performance will be much worse. If you want to switch the operating style, you must retrain AlphaStar from the beginning. In short, the system lacks sufficient flexibility. This feature will quickly amplify the training cost of DeepMind. The data will not lie. According to the latest financial report submitted by DeepMind to the British Companies Registry at the end of last year. The report shows that in the past three years, DeepMind’s losses were 477 million pounds (2019); 470.2 million pounds (2018); 302 million pounds R (2017). In terms of revenue, DeepMind’s revenue in 2019 reached 266 million pounds; in 2018 it reached 103 million pounds; 2017 revenue was 54.42 million pounds. Based on the performance of the past three years, it is not difficult to find that DeepMind’s revenue has indeed grown steadily, and its loss exposure has gradually decreased. However, it should be pointed out that most of DeepMind’s customers rely on affiliates of its parent company Alphabet. To a certain extent, this also shows that DeepMind’s commercialization still has very high room for improvement. This is true for leading AI companies, and domestic AI game makers are estimated to be pretty much the same. Only by effectively solving the problems of sample, cost, flexibility, etc., can AI games really explode