Tencent Reveals How ‘King’s Glory’ AI Beats Professional Stakes

In the ad hoc environment of the semi-finals of the World Champions Cup in Kuala Lumpur in August this year, Tencent AI Lab and Tianmei Studios developed a strategic collaborative AI “Disillusionment” to win the 5V5 match with professional players, while the same year in Shanghai ChinaJoy Live, “The Vision”, “It also opened up the experience test against the amateur 1V1, with a total of 2,100 fights over four days, resulting in a 99.8% win.”

Tencent Reveals How 'King's Glory' AI Beats Professional Stakes

Tencent AI Lab and Tianmei Studios recently published a joint paper detailing the development story behind the undefeated record of “The Disillusion.”

As the researchers point out in their paper, real-time strategic MOBA games such as King’s Glory are different from traditional board games and Atari games, which require more difficult and complex operations and are more complex in their playing environment. In the case of “Glory of kings”, a game may involve 10 to 600 game possibilities and 10 s 18000 possible game operations, and this is only the basis, AI also need to launch GANK, defense, induction opponents, re-upts, etc. in the MOBA game, and may also involve complex skill series.

Tencent Reveals How 'King's Glory' AI Beats Professional Stakes

Tencent researchers systematically encode image features and game status information to digitally represent different units and enemy targets in the game, which runs on a total of 600,000 processors and 1064 graphics cards, including NVIDIA Tesla P40s and NVIDIA V100s. 16,000 contain non-hidden unit attributes and game information. Training a hero requires 48 graphics cards and 18,000 processor cores, and the system continues to train at 80,000 samples per second. The current level of training in the system is equivalent to the 500 years of human experience gained.

Tencent Reveals How 'King's Glory' AI Beats Professional Stakes

According to the researchers, “disillusionment” has been “out of the division” after 80 hours of systematic training, but after 30 hours of training, “disillusionment” is already capable of beating the top 1% of players. Fully trained AI will be able to react in 0.1 seconds, which is already equivalent to the reaction of a top amateur.

In the fight with real players, although the opponent is an experienced professional player, but “excellent” still achieved an average of five kills, while the field average of only 1.33 data. At this year’s ChinaJoy event, “Fantastic” played 2100 1V1 matches with live players, with a winning rate of 99.81 percent, of which five of the eight Heroes controlled by AI achieved a 100% winning percentage.

Tencent researchers say they plan to open their frameworks and algorithms in the near future to promote research into complex games such as King’s Glory.

Add a Comment

Your email address will not be published. Required fields are marked *