PaperlayerView exact revision on arXiv

Source-linked AI summary

CHAMP: Cross-domain Hybrid Architecture for Matchmaking and Prediction in Online Multi-Player Games

Kai Wang, Ge Fan, Chaoyun Zhang, Yuyang Jiang, Yuze Liu

arXiv:2609.04870v1cs.AI

TL;DR

CHAMP addresses cold start, distribution inconsistency, and data sparsity in cross-mode MOBA matchmaking. It combines hybrid cross-mode features with the domain-aware DAWN network and serves all supported modes through one model. DAWN reaches 67.73% offline accuracy, while online tests reduce 5-minute kill crushing by up to 20.73% for novice players.

  • Problem

    Single-mode matchmaking predictors face cold-start, distribution-inconsistency, and data-sparsity challenges across modes and heterogeneous player populations.

  • Method

    CHAMP combines cross-mode sequences, per-mode statistics and real-time signals with DAWN, which learns mode-conditioned representations and per-mode debiasing in one shared network.

  • Results

    67.73% win-rate prediction accuracy and up to 20.73% lower 5-minute kill crushing for novice players are reported for CHAMP.

  • Takeaways & Limitations

    A single domain-aware DAWN model can serve multiple MOBA modes while supporting online reductions in imbalanced matches across League tiers.

  • Takeaways & Limitations

    The empirical setup treats game modes and rank tiers as domains, with offline experiments aggregating modes and online evaluation stratifying by rank tier.

Abstract

from arXiv · show

Multiplayer Online Battle Arena (MOBA) games rely on matchmaking to maintain competitive balance. Our prior work, CUPID, framed matchmaking as an assignment re-optimization problem and showed that a single-mode win-rate predictor can meaningfully rebalance teams. However, deploying such a system across diverse player populations exposes three practical bottlenecks: most queueing players lack sufficient in-mode match history (cold start), skill distributions shift drastically across rank tiers (distribution inconsistency), and extreme skill segments are severely data-starved. We present CHAMP, a cross-domain matchmaking framework that resolves these deployment bottlenecks. To address data sparsity and cold starts, CHAMP replaces the target-mode-only player profile with a hybrid domain feature collection: a timestamp-ordered cross-mode short-term sequence whose slices are annotated with target-domain features, plus per-mode breakdowns of long-term, real-time and team statistics. We further propose the Domain-Aware Win-rate Network (DAWN): a Domain-aware Knowledge Extractor (DAKE) compiles target-mode attributes into learnable representations that feed Domain-Aware Temporal/Spatial/Permutation OmniNet Encoders (DATOE/DASOE/DAPOE), so that mode-conditioned representations and per-mode debiasing are learned jointly inside a single shared network. Online, one trained DAWN serves every supported mode, with per-mode position-satisfaction thresholds as the only mode-specific knob. Offline, DAWN achieves 67.73% win-rate prediction accuracy, outperforming all evaluated attention and sequence baselines. Online A/B tests across the entire League ladder of a large-scale MOBA game, from novice players up to the top-expert players served by Elite Mode, demonstrate consistent drops in imbalanced matches. For lower-tier players, CHAMP reduces the 5-minute kill crushing rate by up to 20.73%.

1 Introduction

CHAMP extends Cupid’s single-mode re-matchmaking into a unified cross-mode system targeting cold start, distribution inconsistency, and data sparsity. It combines hybrid player features with DAWN and serves all supported modes through one model, improving offline prediction and online match balance.

  • Motivation: Cupid improves fairness by re-ranking team-and-position permutations, but its win-rate predictor uses features exclusively from one game mode.This single-mode design creates operational and calibration challenges when players and data span multiple modes.
  • Challenges: CHAMP addresses player cold start, distribution inconsistency, and data sparsity across heterogeneous MOBA modes and skill populations.The challenges include limited target-mode history, mode-specific behavior, and especially sparse samples in Elite Mode.
  • Feature design: CHAMP replaces single-mode inputs with cross-mode behavior sequences plus per-mode long-term statistics and real-time signals.The resulting representation describes player state across modes while remaining mode-aware.
  • Model architecture: DAWN extends the OwO backbone with DAKE and three Domain-Aware encoders that jointly learn mode-conditioned representations and per-mode debiasing.The encoders are DATOE, DASOE, and DAPOE; DAWN achieves 67.73% prediction accuracy and surpasses evaluated baselines, including Cupid’s OwO model.
  • Deployment and results: Online, one DAWN model serves all supported modes while per-mode satisfaction thresholds express mode-specific position preferences.A/B tests across League tiers show consistent reductions in economy and kill crushing rates, including a maximum 20.73% reduction in 5-minute kill crushing for the lowest tier.

2 Related Work

Related work spans scalar skill estimation, win prediction, re-matchmaking, and domain-aware modeling. CHAMP applies multi-domain conditioning to MOBA win prediction by injecting target-mode context into the shared OwO architecture.

  • Matchmaking and skill estimation: Traditional MOBA matchmaking uses scalar MMR, while later systems model player interactions, optimize global match quality, or improve cold-start skill estimation.Examples include ELO, TrueSkill, OptMatch, GloMatch, and QuickSkill.
  • Win prediction and re-matchmaking: Win prediction models forecast outcomes either before matches from rosters and history or during matches from live game state.Cupid bridges prediction and matchmaking by re-optimizing team and position assignments with the OwO encoder.
  • Domain-aware modeling: Domain-aware research includes cross-domain transfer learning and multi-domain architectures that share parameters while retaining domain-specific structure.CHAMP adopts this paradigm by treating each game mode as a domain.
  • CHAMP’s position: CHAMP injects learnable target-mode representations into the OwO backbone through Domain-Aware encoders.This supports mode-conditioned representations and per-mode debiasing within one shared network rather than a post-hoc reweighting layer.

3 Preliminary

The preliminary framework separates scalable lobby formation from fine-grained re-matchmaking, which optimizes team and position assignments inside each lobby. CHAMP treats heterogeneous modes and rank populations as domains for cross-domain prediction.

  • Pre-matching: Pre-matching selects ten players within a narrow MMR window, prioritizing approximate skill parity and short queue times at production scale.Its scale requirements favor inexpensive scalar MMR-window heuristics over per-lobby modeling.
  • Re-matchmaking: Re-matchmaking searches the smaller intra-lobby assignment space, where position errors and player synergies can still produce blowouts despite balanced MMR.This stage can spend modeling capacity on team-and-position permutations because it operates within each grouped lobby.
  • Candidate generation: Candidate generation enumerates team-and-position assignments and prunes those whose aggregate position satisfaction falls below threshold Pτ.The surviving assignments are intended to have acceptable role mappings for all ten players.
  • Final selection: Final selection predicts each surviving candidate’s win rate, converts it to a fairness score, and chooses the assignment closest to a 50% predicted matchup.The fairness score peaks at 1 when predicted win rate equals 0.5 and declines for more lopsided outcomes.
  • Prediction constraint: Win-rate prediction is the binding constraint on online fairness because miscalibration can select lopsided matches when fair candidates exist.This motivates the paper’s focus on cross-domain win-rate prediction.
  • Cross-domain setup: CHAMP treats each game mode or rank-tier population as a separate domain, aggregating modes offline while stratifying online evaluation by rank tier.For a request in one domain, the system can use limited in-domain data together with knowledge from other domains.

4 The Design of CHAMP

CHAMP redesigns Cupid into a unified cross-mode re-matchmaking system through hybrid player features, domain-aware win-rate prediction, and shared online deployment across game modes.

  • CHAMP retains Cupid’s re-matchmaking loop while redesigning feature collection, win-rate prediction, and deployment for all supported game modes.
  • Hybrid Domain Feature Collection: Hybrid Domain Feature Collection combines a timestamp-ordered cross-mode short-term sequence with target-domain annotations and per-mode long-term, real-time, and team statistics.The resulting representation provides a richer, multi-resolution input than a target-mode-only profile.
  • DAWN Architecture: DAWN uses DAKE to encode target-mode attributes and feeds the resulting domain representation into the DATOE, DASOE, and DAPOE encoders.The encoders jointly learn mode-conditioned player representations and match-level context within the shared network.
  • Unified Online Deployment: A single trained DAWN instance serves every supported mode, with per-mode position-satisfaction thresholds as the only mode-specific deployment setting.Adding a mode requires registering its attribute vector and threshold rather than training or hosting a separate model.
  • DAWN Architecture: DAPOE’s pooled domain context supplies match-level information and an in-network debiasing channel, keeping shared player representations more transferable across modes.
  • Unified Online Deployment: Online processing retrieves hybrid features, filters assignments by position satisfaction, scores candidates with DAWN, and returns the assignment whose predicted win probability is closest to 50%.

5 Experiments

The experiments evaluate DAWN against baselines, ablate its components, and test CHAMP online across Casual, League, and Elite modes. Offline evaluation uses balanced win prediction with ACC and RMSE, while online testing measures match-imbalance outcomes.

  • The study answers three questions covering baseline comparison, component ablations, and live matchmaking effects.RQ1 and RQ2 are evaluated offline; RQ3 is evaluated through large-scale online A/B testing.
  • The datasets contain approximately 40M Casual, 30M League, and 0.1M Elite samples, making Elite the primary sparsity stress test.Elite serves the smallest top-expert player pool and has orders of magnitude fewer samples than the other modes.
  • DAWN is compared with LR, MLP, LSTM, Transformer, OwO, and Dawn-single to isolate the effect of joint multi-mode training.Dawn-single uses the same single-target-mode regime as the baselines.
  • Accuracy measures discrete classification quality, while RMSE measures calibration of predicted win probabilities; higher ACC and lower RMSE are better.The task is naturally class-balanced because each match has one winning and one losing team.
  • Ablation experiments remove the hybrid feature collection, DAKE, or each Domain-Aware encoder while keeping the remaining architecture intact.Figure 2 reports these ablations on Elite, with Casual and League showing the same monotone ordering.

5.2 Results Compared with Baselines (RQ1)

DAWN consistently outperforms the evaluated baselines across Casual, League, and Elite, with its advantage over Dawn-single increasing in the sparsest setting. Elite shows the clearest benefit from cross-mode joint training.

  • DAWN ranks first on both ACC and RMSE across all three datasets, followed by Dawn-single and then the evaluated baselines.The consistent ordering indicates that the gains are not specific to one dataset.
  • Dawn-single barely exceeds OwO on Elite at 0.6580 versus 0.6559 ACC, a +0.32% difference.Dawn-single represents single-target-mode training, matching the regime used by the baselines.
  • 0.6773 ACC is achieved by DAWN on Elite, a +2.93% gain over Dawn-single and its largest mode-wise improvement.The gap between DAWN and Dawn-single widens on the sparsest dataset.

5.3 Ablation Studies (RQ2)

Ablations show that every major DAWN component contributes on the sparse Elite dataset, while removing DAKE causes the most severe degradation. Unconditioned pooling of modes can perform worse than single-mode training.

  • The w/o HF, w/o DAKE, w/o DATOE, w/o DASOE, and w/o DAPOE variants each remove one component while retaining the rest of DAWN.The ablation design isolates the contribution of each feature or encoder component.
  • ACC drops by 0.61–7.34 points when any one of the five components is removed from DAWN on Elite.RMSE follows the same monotone degradation pattern, indicating complementary contributions across the hybrid features, DAKE, and three Domain-Aware encoders.
  • Removing DAKE reduces Elite ACC to 0.6039, below the 0.6559 achieved by single-mode OwO.The w/o DAKE variant is vanilla OwO trained on pooled modes without explicit mode conditioning.

5.4 Online Experiments (RQ3)

Online A/B tests against Cupid show that CHAMP reduces match-crushing rates across League and Elite and across the reported horizons and metrics. Elite generally benefits more than League, especially for early-game economy imbalance.

  • CHAMP reduces 5-minute kill blowouts by 9.00% in League and 9.62% in Elite.The online experiment deployed CHAMP in both production modes and compared it with Cupid as the control.
  • CHAMP reduces 15-minute economy blowouts by 3.45% in League and 3.84% in Elite.Negative shifts indicate fewer matches exceeding the critical economy-differential threshold.
  • The reductions have consistent signs across all ten mode-by-metric cells, rather than appearing only at one horizon or for one metric.The reported indicators cover economy and kill crushing rates across the deployed modes and time horizons.
  • Elite shows a 55% larger reduction than League in 5-minute economy crushing, at −3.32% versus −2.14%.The paper relates Elite’s stronger lift to its sparse data regime and narrow top-expert skill band.

6 Lessons Learned

CHAMP’s deployment lessons center on conditioned cross-mode transfer, cohort-specific evaluation, and a reduced operational surface. Its production gains concentrate among cold-start players, while shared serving keeps mode expansion maintainable.

  • Cross-Mode Data Is Not Free Without Conditioning: Naively pooling modes without domain conditioning lowers Elite accuracy to 0.6039, versus 0.6559 for an Elite-only OwO model.The result shows that cross-mode supervision requires explicit conditioning to handle heterogeneous distributions.
  • Cold-Start Cohorts Drive the Production Lift: Cohort-by-cohort evaluation is necessary because aggregate lift can understate gains on the cold-start slice.Table 3 reveals an asymmetry that is not visible in aggregate League results.
  • Cold-Start Cohorts Drive the Production Lift: 20.73% reduction in 5-minute kill crushing occurs for Novice players, compared with −9.00% across League and −6.30% for Expert players.The reduction shrinks monotonically from Novice through Expert, concentrating the production payoff in thinner-history cohorts.
  • Cold-Start Cohorts Drive the Production Lift: CHAMP’s cross-domain transfer supplies missing target-mode signal where in-tier behavioral histories are thinnest.The paper interprets the Novice gain as imputing missing target-mode information from cross-mode behavior.
  • Collapse the Mode-Specific Operational Surface: CHAMP reduces mode-specific deployment configuration to one position-satisfaction threshold, τ_m, while shared parameters absorb other mode dependence.The design replaces N per-mode deployment artifacts with one shared serving system.
  • Collapse the Mode-Specific Operational Surface: Opening Elite Mode shifted QPS between modes while leaving global QPS essentially unchanged.Elite QPS rose as League QPS fell comparably, whereas independent per-mode services would require capacity planning for separate peaks.

7 Conclusion

CHAMP is a unified cross-domain re-matchmaking framework using DAWN to serve multiple MOBA modes through one shared model. It achieves 67.73% offline accuracy and reduces 5-minute kill crushing by up to 20.73% for novice players online.

  • Unified Framework: CHAMP serves multiple MOBA game modes through a single shared DAWN model.Its unified framework addresses cold-start and data-sparsity bottlenecks through joint representation learning with DAKE.
  • Empirical Results: 67.73% offline accuracy is achieved for win-rate prediction.The reported operational guidelines also require domain-aware conditioning when using cross-mode data.
  • Empirical Results: 20.73% is the maximum online reduction in 5-minute kill crushing for novice players.The conclusion identifies concentrated cold-start gains and a single-scalar configuration as key deployment guidelines.

GenAI Usage Disclosure

The authors used large language models for manuscript grammar checking and polishing, while the technical content, experiments, and results remain solely their work.

  • GenAI Usage Disclosure: Large language models, including ChatGPT, were used for grammar checking and manuscript polishing.The disclosure separates language assistance from the paper’s technical work.
  • GenAI Usage Disclosure: The authors state that all technical content, experiments, and results are solely their work.
Loading 2609.04870v1…