Research map / Learning theory and algorithms / Submodular and game algorithms
EnMatch: Matchmaking for Better Player Engagement via Neural Combinatorial Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EnMatch, an engagement-oriented k-vs-k matchmaking framework that couples a learned engagement model with a PPO-trained graph-encoder/masked-decoder neural combinatorial optimization policy and an LNS heuristic operator.
Problem
Tier-based fair matchmaking ignores player engagement; engagement-oriented team formation in k-vs-k games is an NP-hard combinatorial optimization with non-linear (neural) objectives and large search spaces.
Approach
Formulate matchmaking as an MDP where a neural engagement model provides rewards, and a GAT encoder with a Transformer masked decoder (KNN-based pointer-like selection) generates 2K players per match, refined by a large-neighborhood-search heuristic operator and trained with PPO.
Key results
- Engagement model EnMatchNet reaches ACC 0.734 / AUC 0.683 on SPG and 0.798 / 0.774 on RPGPVP, 2.6-3.2% ACC gains over the second-best method.
- In simulation, EnMatch improves average reward over the best baseline by 2.15% on SPG (66.5 vs 65.1) and 11.76% on RPGPVP (58.9 vs 52.7).
- In a two-week online A/B test, EnMatch increases total matches played by 1.64% on SPG and 7.19% on RPGPVP over the best baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the auxiliary task during engagement-model training (ablation) | EnMatchNet · SPG | ACC: 0.708 -> 0.734 | Helped |
| add the auxiliary task during engagement-model training (ablation) | EnMatchNet · SPG | AUC: 0.664 -> 0.683 | Helped |
| add the auxiliary task during engagement-model training (ablation) | EnMatchNet · RPGPVP | ACC: 0.778 -> 0.798 | Helped |
| add the auxiliary task during engagement-model training (ablation) | EnMatchNet · RPGPVP | AUC: 0.751 -> 0.774 | Helped |
| use EnMatch instead of the strongest baseline | matchmaking · SPG simulation | cumulative reward per episode: 2.15% | Helped |
| use EnMatch instead of the strongest baseline | matchmaking · RPGPVP simulation | cumulative reward per episode: 11.76% | Helped |
| use EnMatch instead of the strongest baseline | online matchmaking · SPG online deployment | number of matches: 1.64% | Helped |
| use EnMatch instead of the strongest baseline | online matchmaking · RPGPVP online deployment | number of matches: 7.19% | Helped |
Surprising findings
- Fair-game matchmaking is not sufficient for engagement: cross-tier 'diversity teams' showed higher chat and upvote rates and higher retention than fair teams in the data analysis.
- Main players with three consecutive wins did not have clearly lower churn than other scenarios, suggesting diverse win-loss and role experiences boost engagement.
Limitations
- Evaluated on proprietary industrial game datasets and a deployment on two games, limiting reproducibility.
- Simulation relies on a learned engagement model as the reward rather than real player-level feedback for training.
Experimental setup
Datasets: SPG RPGPVP
Benchmarks: SPG RPGPVP online A/B test on two games
Metrics: accuracy AUC average cumulative reward per episode number of matches played
Baselines: Tier-based Heuristics PointerNetwork GloMatch OptMatch EOMM MBNet Random
Builds on: Proximal Policy Optimization Pointer Network EOMM GloMatch Large Neighborhood Search
Closest papers by meaning
- Learning from Group Comparisons: Exploiting Higher Order Interactions (NeurIPS 2018 · similarity 0.86)
- From Average-Iterate to Last-Iterate Convergence in Games: A Reduction and Its Applications (NeurIPS 2025 · similarity 0.83)
- Online Matching with Stochastic Rewards: Provable Better Bound via Adversarial Reinforcement Learning (ICML 2024 · similarity 0.83)
- The CoachAI Badminton Environment: A Novel Reinforcement Learning Environment with Realistic Opponents (Student Abstract) (AAAI 2024 · similarity 0.83)
- Learning Diverse Risk Preferences in Population-Based Self-Play (AAAI 2024 · similarity 0.83)
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.