How Self‑Play and GAIL Powered the WeKick AI to Win the First Google Football Kaggle Championship

During a relaxed New Year break, the author enjoyed a game of King of Glory with friends, recalling the intense challenge mode featuring Tencent's powerful computer AI, "Juewu" (Absolute Insight), which proved far tougher than ordinary bots.

After struggling against Juewu, the author finally succeeded using a "Da Qiao‑Mylady" teleport‑steal strategy.

Suddenly, news broke that China had won a football World Cup—not a traditional men's or women's tournament, but the AI‑driven "WeKick" team, which claimed the title in the inaugural Google Football Kaggle competition, outperforming 1,138 other teams with a score of 1,785.8.

The article explains why controlling a football AI is far more demanding than a MOBA: the number of agents rises from 5v5 to 11v11, requiring long‑term planning, rapid decision‑making, and coordination of speed, acceleration, shooting, heading, passing, and defense for each player.

Self‑Play Reinforcement Learning Framework

WeKick employed a self‑play reinforcement learning approach, training models from scratch within an asynchronous distributed framework. Although this sacrifices some real‑time performance, it offers flexibility to scale resources and adapt to the larger number of agents in football.

GAIL Generative Adversarial Imitation Learning

To bridge the gap between MOBA and football, the team combined GAIL with handcrafted rewards, allowing the AI to learn expert behavior from other teams and use the GAIL‑trained model as a fixed opponent for further self‑play refinement.

League Multi‑Style Reinforcement Learning

The self‑play method tended to produce a single playing style, so WeKick introduced a League of multiple strategy pools to diversify tactics. The training pipeline follows three steps:

Train base models for specific skills such as dribbling, passing, shooting, and tackling.

Generate several style‑specific models, each focusing on a distinct play style, using the main model as an opponent to avoid rigidity.

Train a master model that pits its historical versions and all style models against each other, ensuring robust responses to any opponent.

This approach boosted the master model’s internal rating by 200 points over the best style model and added another 80 points.

The Kaggle competition, launched in 2021, was the first football‑AI challenge on the platform. The exponential growth in difficulty from controlling 5 agents to 11 agents makes the problem a benchmark for cutting‑edge multi‑agent reinforcement learning.

Previously, Tencent’s AI "Juewu" had already won the Google Research Football League (a 5v5 competition). The evolution from Go AI "Jueyi" to MOBA AI "Juewu" and now to football AI "WeKick" demonstrates Tencent’s advancing deep reinforcement learning capabilities, hinting at broader future applications.

The author ends with a wish to face the WeKick AI in a match, inviting readers to imagine the challenge.