SenseTime’s AI Agent DI-star Reaches Parity with Professional Gamers and Pushes New Boundaries for AI Decision Making

June 13, 2022 – SenseTime recently demonstrated its AI agent DI-star, the most powerful opensourced AI model specifically developed for the real-time strategy (RTS) game “StarCraft II”. Demonstrated publicly for the first time, it successfully reached parity with top professional players in multiple games, making a breakthrough in the application of AI decision-making in video games.

Zhou Hang, an 8-time championship of StarCraft II in China, now working as the Head of Technical Research and Development in the AI Decision and Game Division at SenseTime, said, “DI-star’s performance levels are comparable to professional players only after five weeks of training. Such efficient training results are the result of SenseTime’s leading strength in AI decision-making and the powerful computing support provided by its proprietary AI infrastructure SenseCore.”

DI-star and the decision-making AI platform OpenDILab have both been opensourced on GitHub to promote large-scale application of AI technology across the video game industry, as well as create an AI innovation ecosystem for video games.

DI-star’s performance levels are comparable to professional players. 

Accurate Decision-making and High-performance

In recent years, AI has demonstrated its ability to defeat humans in chess, Go and various computer games. "StarCraft II" requires strong predictive ability, cognitive reasoning and fuzzy decision-making capabilities. With its full-stack AI capabilities in decision intelligence, SenseTime fully demonstrated DI-star's flexible decision-making ability in this acclaimed RTS game, which can quickly find the best strategy for each game.

DI-star allows the AI agent to adopt a self-gaming approach and conduct a large number of games simultaneously. Combining cutting-edge technologies like supervised learning and reinforcement learning, DI-star continues to evolve through self-confrontation, finally achieving a competitive level that is comparable to top-ranked human players.

Fully Supported by SenseCore’s Capabilities

Leveraging high-performance algorithms and the excellent computing power of SenseCore, which provides a solid foundation for model building, training and verification, DI-star managed to complete 100 million games in just five weeks. SenseCore also provides the necessary production tools and deployment tools for DI-star to use extensive trials and error in training, driving the algorithms to iterate at high speed.

Meanwhile, to empower the development of the video game industry, SenseTime has created various solutions, such as AI Bot, intelligent creation of game content and game art, with its comprehensive AI technology capability. 

Liu Yu, General Manager of AI Decision and Game Division at SenseTime, said, “In the future, SenseTime will leverage SenseCore’s capabilities to promote diverse applications of AI in video games and join industry partners to encourage innovation and development of the video game industry.”