Introduction
In traditional sports, the scoreboard provides a clear, real-time metric of team performance. However, in Multiplayer Online Battle Arenas (MOBAs) like DotA2 or League of Legends, the "score" is multifaceted and abstract. A team trailing in kills might hold a significant gold advantage or possess a "late-game" hero composition poised for a comeback. Evaluating the real-time state of a MOBA match is notoriously difficult due to the massive state space, including hundreds of heroes, thousands of item combinations, and complex spatial dynamics.
Quantifying this "relative advantage" is essential for developing advanced gaming AI, providing reinforcement learning reward functions, and enhancing spectator analytics. The MOBA-Slice framework addresses this by treating the game result as a time-dependent discounted reward.
Methodology: The Time Slice Evaluation (TSE) Model
The core of the MOBA-Slice framework is the Time Slice Evaluation (TSE) model, which moves beyond binary win/loss predictions to provide a continuous advantage score.
1. Discounted Evaluation Function
Inspired by reinforcement learning, the authors define advantage as a function of remaining time. The intuition is that a lead is more "crushing" if it leads to a quick victory. The formula $DETS(R, t) = R / (\alpha t)$ (where $R$ is the result and $t$ is remaining time) creates a target value for the neural network, allowing it to learn the intensity of an advantage.
2. Hybrid Neural Architecture
The TSE model employs a hybrid architecture to capture both individual performance and global team dynamics:
- Individual (Ind) Part: This component uses specialized sub-networks for each hero (over 114 in DotA2). It learns the specific contribution of a hero based on their individual stats like gold, experience, and items.
- Global (Glo) Part: A larger multilayer perceptron processes the entire match state to capture synergies, hero counters, and team-wide positioning.
- Fusion Layer: The outputs from the Individual and Global parts are fused to produce a final relative advantage score ranging from -1 to 1.
Dataset and Features
The model was trained on a massive dataset of 71,355 professional DotA2 matches, comprising approximately 2.8 million individual time slices sampled every 60 seconds. Features extracted for each slice include:
- Hero Attributes: HP, gold, experience, and map coordinates.
- Statistics: Kills, deaths, and tower damage.
- Items: A vector representing which of the 244 possible items each hero possesses.
Key Results and Insights
The MOBA-Slice framework demonstrates significant improvements over existing baselines:
- Superior Accuracy: The model achieved a win-prediction accuracy of 71.9%, outperforming Valve’s commercial "DotA Plus Assistant" by 3.7%.
- Time Estimation: The model can estimate the remaining match duration with a mean absolute error (MAE) of 7.7 minutes during the latter half of the game.
- The Importance of Synergy: Ablation studies revealed that the "Global" part of the model is more influential than the "Individual" part, confirming that team composition and interaction are more critical than raw individual stats in high-level play.
- Stability in the Late Game: The researchers found that the relationship between game state and outcome stabilizes in the second half of matches, making predictions more reliable as the game progresses.
Limitations and Future Work
While effective, the current model has limitations. It currently omits hero skill levels and the specific health of non-hero units (creeps and buildings). Furthermore, the model assumes a monotonic growth of advantage, which may not capture sudden "comebacks" or "throws" effectively. Future iterations could incorporate Recurrent Neural Networks (RNNs) or LSTMs to model the temporal "momentum" of a match rather than treating each time slice in isolation.
Summary updated