[Paper Review] Which Heroes to Pick? Learning to Draft in MOBA Games with Neural Networks and Tree Search

Sheng Chen, Menghui Zhu+4 more|arXiv (Cornell University)|Dec 18, 2020

Artificial Intelligence in Games41 references4 citations

TL;DR

This paper proposes JueWuDraft, a novel hero drafting algorithm for MOBA games that combines neural networks with Monte-Carlo tree search to optimize multi-round (best-of-N) drafting. By modeling the drafting process as a combinatorial game and introducing a long-term value estimation mechanism, JueWuDraft achieves superior performance over state-of-the-art methods, particularly in balancing short-term gains with long-term win probability across multiple rounds.

ABSTRACT

Hero drafting is essential in MOBA game playing as it builds the team of each side and directly affects the match outcome. State-of-the-art drafting methods fail to consider: 1) drafting efficiency when the hero pool is expanded; 2) the multi-round nature of a MOBA 5v5 match series, i.e., two teams play best-of-N and the same hero is only allowed to be drafted once throughout the series. In this paper, we formulate the drafting process as a multi-round combinatorial game and propose a novel drafting algorithm based on neural networks and Monte-Carlo tree search, named JueWuDraft. Specifically, we design a long-term value estimation mechanism to handle the best-of-N drafting case. Taking Honor of Kings, one of the most popular MOBA games at present, as a running case, we demonstrate the practicality and effectiveness of JueWuDraft when compared to state-of-the-art drafting methods.

Motivation & Objective

• To address the limitations of existing MOBA drafting methods that fail to handle large hero pools and multi-round match structures.
• To model the best-of-N drafting process as a multi-round combinatorial game with long-term strategic value.
• To design a value estimation mechanism that prioritizes future round outcomes rather than immediate win probability.
• To improve drafting efficiency and effectiveness in large-scale MOBA games like Honor of Kings with over 100 heroes.
• To develop a practical, scalable AI drafting system applicable to real-world MOBA gameplay and human player assistance.

Proposed method

• Formulate the MOBA drafting process as a multi-round combinatorial game, modeling team selection across multiple matches.
• Integrate a neural network-based winning rate predictor to estimate the outcome of complete lineups.
• Apply Monte-Carlo Tree Search (MCTS) with a long-term value function to guide draft decisions beyond immediate gains.
• Design a value mechanism that maximizes the mean predicted winning rate across all rounds in a best-of-N series.
• Use simulated rollouts with the neural network to estimate value, avoiding exhaustive search in large hero pools.
• Implement a non-greedy strategy that considers future draft states and opponent counterplay, improving long-term win probability.

Experimental results

Research questions

• RQ1 How can hero drafting in MOBA games be effectively modeled as a multi-round combinatorial game?
• RQ2 Can a neural network-based value estimator improve drafting decisions beyond single-round optimization?
• RQ3 Does a long-term value mechanism in MCTS lead to better performance in best-of-N MOBA series compared to greedy or single-round strategies?
• RQ4 How does the proposed method scale to large hero pools (e.g., 100+ heroes) while maintaining computational efficiency?
• RQ5 To what extent does the algorithm outperform existing methods like OpenAI Five and DraftArtist in real-world MOBA settings?

Key findings

• JueWuDraft outperforms state-of-the-art methods, including OpenAI Five and DraftArtist, in all experimental settings on Honor of Kings.
• The algorithm demonstrates superior efficiency and effectiveness, particularly in handling large hero pools and complex draft sequences.
• JueWuDraft avoids greedy choices—such as selecting the hero with the highest immediate predicted win rate—by prioritizing long-term strategic balance.
• In best-of-3 series, JueWuDraft selects heroes that yield higher predicted winning rates in the second and third rounds, even if the first-round win rate is slightly lower.
• The long-term value mechanism leads to more stable and consistent performance across multiple rounds, reducing reliance on early-game advantages.
• The method has been deployed in real-world applications, including PVE game modes and human player hero recommendations in Honor of Kings.