New MOBA-Playing AI System Can Beat 99.81% of Human Opponents
Thu, April 22, 2021

New MOBA-Playing AI System Can Beat 99.81% of Human Opponents

AI can beat human opponents in multiplayer online battle arena games (MOBAs), particularly the “Honor of Kings" / Credits: Gorodenkoff via Shutterstock

 

Artificial intelligence has now entered the world of online games. A recent study conducted by Tencent, a Chinese multinational conglomerate holding company, introduced a new technique to explore the game maps of multiplayer online battle arena games (MOBAs), particularly the “Honor of Kings,” efficiently. The researchers chose real-time strategy games like Honor of Kings because they are much more complex than traditional board games and Atari games. 

In real-time strategy games, players need to learn how to plan, attack, and defend. At the same time, they need to control skill combos, induce, and deceive opponents while contending with hazards like creeps and fully automated turrets. Tencent developed an AI system capable of defeating teams in real-time strategy games Honor of Kings that can beat 99.81% of its human opponents. This aims not to achieve Honor of King's superhero performance, but to develop systems capable of solving some of society’s toughest challenges.

According to VentureBeat, an American technology website that publishes news, analysis, long-form features, interviews, and videos, the AI system achieved five kills per game and was killed only 1.33 times per game on average. This is equivalent to a win rate of 99.81% over 2,100 matches, while five of the eight AI-controlled heroes managed a 100% win rate. The AI system is currently at the top of beating human players, replacing DeepMind’s AlphaStar beating 99.8% of human StarCraft 2 players. 

Also, Tencent’s architecture consists of four modules. This includes an AI server which dictates how the AI model interacts with objects in the game environment; a Dispatch Module which collects data samples consisting of rewards, features, action probabilities; a Memory Pool which supports samples of various lengths and data sampling based on the generated time, and a Reinforcement Learner which accelerates policy updates. In the future, the researchers are planning to make both their framework and algorithms open source.