CHAMP:在线多人游戏中用于配对与预测的跨域混合架构
文章背景与核心概要
多人在线战术竞技(MOBA)游戏极大地依赖于匹配系统来确保公平和具有竞争性的对局。尽管以往的方法(如 CUPID)通过单一模式的胜率预测器,将匹配重新构建为分配再优化问题,但它们在现实世界中面临着严峻的瓶颈:冷启动(玩家在特定模式中缺乏足够的对局历史)、分段间的分布不一致以及极端分段的数据稀疏。
为了克服这些挑战,作者引入了 CHAMP 这一跨域匹配框架。CHAMP 利用混合特征采集和域感知胜率网络(DAWN),统一了跨多个模式的玩家画像。离线评估表明,DAWN 实现了 67.73% 的胜率预测准确率,而大规模在线 A/B 测试则证明,它显著减少了不平衡对局,并将新手玩家的游戏前期“碾压”率降低了 20.73%。
arXiv ID: arXiv:2609.04870 [cs.AI]
Accepted at: CIKM 2026 (Applied Research Track)
Submission Date: September 4, 2026
Authors: Kai Wang, Ge Fan, Chaoyun Zhang, Yuyang Jiang, Yuze Liu
📌 执行摘要 (Executive Summary)
多人在线战术竞技(MOBA)游戏极大地依赖于匹配系统来确保公平和具有竞争性的对局。尽管以往的方法(如 CUPID)通过单一模式的胜率预测器,将匹配重新构建为分配再优化问题,但它们在现实世界中面临着严峻的瓶颈:冷启动(玩家在特定模式中缺乏足够的对局历史)、分段间的分布不一致以及极端分段的数据稀疏。
Multiplayer Online Battle Arena (MOBA) games rely heavily on matchmaking to ensure fair and competitive matches. While previous approaches like CUPID reframed matchmaking as an assignment re-optimization problem using single-mode win-rate predictors, they faced critical real-world bottlenecks: cold starts (players lacking sufficient match history in a specific mode), distribution inconsistencies across skill tiers, and data sparsity among extreme skill segments.
为了克服这些挑战,作者引入了 CHAMP 这一跨域匹配框架。CHAMP 利用混合特征采集和域感知胜率网络(DAWN),统一了跨多个模式的玩家画像。离线评估表明,DAWN 实现了 67.73% 的胜率预测准确率,而大规模在线 A/B 测试则证明,它显著减少了不平衡对局,并将新手玩家的游戏前期“碾压”率降低了 20.73%。
To overcome these challenges, the authors introduce CHAMP, a cross-domain matchmaking framework. CHAMP leverages a hybrid feature collection and the Domain-Aware Win-rate Network (DAWN) to unify player profiles across multiple modes. Offline evaluations show DAWN achieving a 67.73% win-rate prediction accuracy, while large-scale online A/B tests demonstrate significant reductions in unbalanced matches and a 20.73% reduction in early-game "crushing" rates for novice players.
🛠️ 现有匹配系统中的部署瓶颈 (Deployment Bottlenecks in Existing Matchmaking Systems)
在不同的玩家群体中部署单模式匹配再平衡预测器会暴露出三个实际局限性: 1. 冷启动: 大多数排队玩家缺乏足够的同模式对局历史来准确评估其当前水平。 2. 分布不一致: 技能分布在不同的排位区间发生了巨大变化。 3. 数据荒漠: 极端的技能分段(例如顶级精英玩家)遭遇严重的数据稀疏问题。
Deploying single-mode match rebalance predictors across diverse player populations exposes three practical limitations: 1. Cold Start: Most queueing players lack adequate in-mode match history to accurately gauge their current skill level. 2. Distribution Inconsistency: Skill distributions shift drastically across different rank tiers. 3. Data Starvation: Extreme skill segments (such as top-tier elite players) suffer from severe data sparsity.
🚀 CHAMP 框架与 DAWN 架构 (The CHAMP Framework & DAWN Architecture)
为了解决这些部署瓶颈,CHAMP 引入了两项关键创新:
To resolve these deployment bottlenecks, CHAMP introduces two key innovations:
1. 混合域特征收集 (Hybrid Domain Feature Collection)
CHAMP 没有仅仅依赖目标模式专属的玩家画像,而是采用了: * 带有目标域特征注释的时间戳排序、跨模式短期序列。 * 长期、实时以及团队统计数据的各模式细分。
Instead of relying solely on target-mode-only player profiles, CHAMP utilizes: * A timestamp-ordered, cross-mode short-term sequence annotated with target-domain features. * Per-mode breakdowns of long-term, real-time, and team statistics.
2. 域感知胜率网络 (DAWN)
DAWN 使用统一架构处理这些混合特征: * DAKE(域感知知识提取器): 将目标模式属性编译为可学习的表示。 * Enclosers(DATOE / DASOE / DAPOE): 域感知时空与排列全网编码器(Domain-Aware Temporal, Spatial, and Permutation OmniNet Encoders),在一个共享网络中联合学习模式条件表示和各模式消偏。
在在线部署期间,单个训练好的 DAWN 可以服务所有支持的模式,并仅将每种模式的位置满意度阈值作为其唯一特定于模式的参数。
DAWN processes these hybrid features using a unified architecture: * DAKE (Domain-aware Knowledge Extractor): Compiles target-mode attributes into learnable representations. * Enclosers (DATOE / DASOE / DAPOE): Domain-Aware Temporal, Spatial, and Permutation OmniNet Encoders that jointly learn mode-conditioned representations and per-mode debiasing within a single shared network.
During online deployment, a single trained DAWN serves every supported mode, using per-mode position-satisfaction thresholds as its only mode-specific parameter.
📊 结果与性能 (Results and Performance)
- 离线性能: DAWN 达到了 67.73% 的胜率预测准确率,超越了所有评估过的基于注意力和序列的基线模型。
- 在线 A/B 测试: 在大型 MOBA 游戏的整个天梯赛中进行了测试(从新手到顶级精英模式玩家不等),CHAMP 实现了:
- 在所有分段中,不平衡对局持续减少。
- 低分段玩家的 5 分钟击杀碾压率降低了 20.73%。
- Offline Performance: DAWN achieves 67.73% win-rate prediction accuracy, outperforming all evaluated attention-based and sequence-based baselines.
- Online A/B Testing: Tested across the entire League ladder of a large-scale MOBA game (ranging from novices to top-tier Elite Mode players), CHAMP delivered:
- Consistent drops in imbalanced matches across all tiers.
- A 20.73% reduction in the 5-minute kill crushing rate for lower-tier players.
🔗 链接与资源 (Links & Resources)
- 查看 PDF: arXiv:2609.04870 PDF
- HTML 版本: arXiv HTML (Experimental)
- DOI: 10.48550/arXiv.2609.04870
- View PDF: arXiv:2609.04870 PDF
- HTML Version: arXiv HTML (Experimental)
- DOI: 10.48550/arXiv.2609.04870