The Heart of the Machine is original
Machine Heart Editorial Department
How NPCs can be as smart as people has always been one of the main demands of game development. But early NPCs were often derided by players as "artificially retarded" because their behavioral paradigms were easily explored.
The ratio of algorithm and resource input to the actual output brought by AI is difficult to estimate before the game is put on the market. As a result, "AI" has not really become widespread in games yet.
Small teams can hardly afford their development costs, and even some large teams worry about poor results and backfire.
There are already game AI players who can "beat" human players
But it is not widespread
Games are the soil that gave birth to many AI technologies. Reinforcement Learning (RL) fits perfectly with games in that it can use virtual environments to train AI agents to actively try to discover which actions will yield the most benefits.
In the history of machine learning, many well-known AI models based on game environments have been born, such as AlphaStar for Playing StarCraft 2, OpenAI Five for Dota 2, And So on. With games as a bridge, new technological breakthroughs in the field of AI are also known to us again and again.
The game AI OpenAI Five once beat the human world champion team.
But these are just experiments, AI has been widely used in games for graphics improvement such as the super-resolution algorithm "DLSS", but has made little contribution to the interaction that can impress players. As one of the three branches of machine learning reinforcement learning, the current application direction is mostly in automatic driving, industrial automation, financial trade and even NLP, although the game is also the main application direction, but it has not really been popularized.
In the field of games, people's demand for technologies such as reinforcement learning is actually very much: in battle games, through AI-customized solutions, the corresponding difficulty can be matched according to the user level, so that the user experience is better. In role-playing, racing games, AI can be used to test and design levels. AI that is high enough can even be used for balance testing.
A common saying among players is that a game's graphics determine whether you'll pay attention to it, and game performance determines whether you like it or not.
And high-IQ AI characters with human professional player capabilities are not only attractive to players, but also great for game developers to recruit massive AI "players" to test game bugs and even help make game levels.
At GTC 2021, developers at game company EA demonstrated reinforcement learning to improve level design.
For game developers, training AI models that can be applied to practice requires enough expertise, but not every game studio understands both game engines and deep learning frameworks. The need for massively parallelized computing power is also an obstacle when building models.
Can we generalize and practicalize highly intelligent AI? In recent times, more and more games are gaining AI blessings.
Address core technical needs
Shorten the distance between developers and AI
In October this year, HUAWEI CLOUD released ModelArts for Games, shortening the distance between game developers and AI technologies.
On this platform, developers can customize capabilities such as reinforcement learning. It is based on ModelArts, an artificial intelligence platform with the largest market share in China, and meets the core needs of game teams for AI from four perspectives: AI development, production deployment, team empowerment, and model monetization.
HUAWEI CLOUD has cooperated with game companies to pre-produce multiple types of game AI competitive algorithms, covering mainstream game types such as MOBA, SPG, FPS, and chess and cards. In the platform, AI robots for different types of games have superb capabilities, and at the same time, they are diverse in style, can show different strategies, and can bring players a more realistic game experience.
Backed by Huawei's AI platform, ModelArts for Games can achieve end-to-end full coverage of workflows: from data processing, obtaining advanced algorithms, large-scale training of models, to on-demand deployment processing and management of models, the full cycle of AI applications can be implemented on one platform.
With ModelArts, AI-trained computing power is no longer an issue, and ModelArts provides heterogeneous computing clusters. Training supports massive CPU and GPU free combination, provides large-scale reinforcement learning training capabilities, supports elastic scaling capacity and fault tolerance mechanism, and facilitates users to use mainstream or customized distributed architectures such as IMPALA and Seed for efficient reinforcement learning training. Taking into account a variety of application scenarios, it greatly improves the efficiency of algorithm development and training. The distributed reinforcement learning framework "Xingtian" cluster communication from Huawei's Noah's Ark Laboratory has been optimized. These features increase training speed by up to 10 times, reduce training costs, and users can start and stop training jobs at any time as needed.
In terms of THE AI inference speed that is directly related to the game experience for players, ModelArts can compress the delay to less than 20ms to achieve real-time response.
For game developers who don't know which algorithm will meet their needs, ModelArts for Games also gives some use cases that integrate the most mature algorithms by game type, many of which have achieved good results in data competitions. For AI novices, HUAWEI CLOUD has also opened more than 10 full-code AI tutorials to meet the needs of customers to cultivate their teams and quickly improve their AI team capabilities.
Researchers at Noah's Ark Lab believe that ModelArts has shown broad application prospects in the direction of diversified generation of game AI, NPC training, and automated testing.
Empower game developers
Bring a better user experience
When reinforcement learning technology really lands, the purpose of AI is not to defeat players, but to accompany more players and bring a better experience. ModelArts for Games has the ability to get the right technology for all game developers, big or small.
The HUAWEI CLOUD AI team has worked with many game developers to implement game AI technology. At present, NetEase Fuxi, Xishanju, and Digital Sky are all cooperating with HUAWEI CLOUD ModelArts for Games, and have achieved phased results.
First of all, HUAWEI CLOUD ModelArts for Games can help developers choose the algorithm that is most suitable for their own game scenarios, and reduce the threshold for developers to apply AI.
In a game that Xishanju is developing, the developers and the HUAWEI CLOUD R&D team have jointly created an AI robot based on reinforcement learning algorithms, which is more powerful than the AI robot in traditional games in terms of battle level and strategy choice. After actual game testing, ai trained by Huawei technology played 4664 rounds with built-in bots, with a 95% win rate. When playing against professional players, the winning percentage can be 50 to 50.
If the game development team has the ability to develop its own AI technology, HUAWEI CLOUD can also provide powerful AI computing power and help them monetize the algorithm model commercially.
NetEase Fuxi has its own artificial intelligence laboratory, which has carried out in-depth cooperation with HUAWEI CLOUD in computing platforms and reinforcement learning.
At the first AI Football World Cup held by Google and Premier League Manchester City clubs on the Kaggle platform, the AI of netease Fuxi reinforcement learning team won the runner-up among thousands of teams, and its AI tactical layout ability close to that of professional coaches made people see the huge potential of "AI + football".
During the game, according to the real-time situation on the field, the AI robot can achieve the popular high-level pressing tactics in the current high-level football league through training, and the attack and defense conversion is also very fast. When targeting different players, AI can reflect different styles of offense and defense, bringing a more realistic experience than ever before.
NetEase Fuxi AI's set-piece playing style also impressed the organizers. Brian Prestidge, director of data insights and decision-making technology at City Football Group, the parent company of Manchester City Club, said that the set-piece strategy adopted by NetEase Fuxi AI is very similar to the set-piece tactics of professional coaches in real-world games.
Now on the ModelArts platform, people can already use HUAWEI CLOUD's large-scale heterogeneous cluster training and online real-time inference to create intelligent NPCs needed for different game scenarios, saving a lot of time required for the regular game development process.
In addition to generating highly intelligent NPCs, such as content generation virtual people, intelligent art, etc. are also the direction of AI applications, and more and more needs have emerged.
As a provider of cloud service "infrastructure", MODELArts for Games of HUAWEI CLOUD can bring greater assistance to game companies in the field of AI capabilities, allowing them to focus on game research and development without being plagued by the threshold of applying AI technology.
Linking upstream and downstream partners in the digital creative industries
Inspire a new future for the industry
Huawei Cloud's gaming AI solution allows digital creative developers to focus on core aspects such as content production without spending time on infrastructure. At the HUAWEI CLOUD & HUAWEI TERMINAL CLOUD Service Innovation Summit 2022 held on December 29, the mini-ideas of game innovation were shared on the spot, and according to the HUAWEI CLOUD 3D Modeling Kit AR Engine, low-code creation and AI intelligent creation were realized, and the physical applications of real scenes reconstructed based on HUAWEI CLOUD were realized.
This is an era of collaborative co-creation, and the granularity, complexity and long cycle of industrial digitalization determine that no single company can succeed alone. Jointly building ecology and integrating into ecology is the only way. Looking at the future, the entire digital creative industry will need more collaborative progress in the upstream and downstream of the industrial chain.
Cloud, AI, and 5G will accelerate the convergence of applications from the end side to the cloud, accelerate streaming, and accelerate real-time interaction; film and television and game production are gradually integrated, and cloud rendering accelerates the development of domestic games to high quality; digital human technology is developing rapidly, automatically building the game world through cloud + AI; and the deep combination of cloud games and live broadcasting makes live interaction bring great changes;
The integrated development of technology and ecology will drive the rapid development of China's domestic games to high quality, and more excellent content will stand out, accelerating self-purification and industrial innovation. In this process, HUAWEI CLOUD is becoming the best partner for many digital creative developers and service providers.