Supervised Learning Achieves Human-Level Performance in MOBA Games: A Case Study of Honor of Kings

We present JueWu-SL, the first supervised-learning-based artificial intelligence (AI) program that achieves human-level performance in playing multiplayer online battle arena (MOBA) games. Unlike prior attempts, we integrate the macro-strategy and the micromanagement of MOBA-game-playing into neural networks in a supervised and end-to-end manner. Tested on Honor of Kings, the most popular MOBA at present, our AI performs competitively at the level of High King players in standard 5v5 games.

I. Introduction

Multiplayer online battle arena (MOBA) games, e.g., Dota, Honor of Kings, and League of Legends, have been considered as an important and suitable testbed for artificial intelligence (AI) research due to their considerable complexity and varied playing mechanics [1]–[4]. The standard game mode of MOBA is 5v5, where two opposing teams of five players each compete against each other. In this mode, each individual in a team has to control the actions of one hero in real time based on both the situation dynamics and the team strategy. During the game, a hero can grow stronger by killing enemy heroes, pushing turrets, killing creeps and monsters, and so on. The goal for players in a team is to destroy their enemy’s main structure while protecting their own. In MOBA, the gameplay is also varied as it involves two factors: macro-strategy, i.e., “where to go” on the game map for a hero, and micromanagement, i.e., “what to do” when the hero reaches the expected location. Despite such complexity and mechanism, the methodological explorations on AI systems for MOBA-game-playing are still very limited [5].