Multimodal Model for Game Frame Rate Prediction

Background

In the game War Intent , players control a general in realistic melee combat, generating complex rendering, effects, and environments that challenge optimization and cause frame‑rate drops, illustrating the need for accurate frame‑rate prediction.

Data and Model

Two data categories affect frame rate: static data (hardware model, graphics settings, resource size) and temporal data (hardware usage logs, player state sequences). Static features can be extracted with a multilayer perceptron, while temporal features use recurrent neural networks; both are fused in a multimodal model.

Multimodal Model

Multimodal learning integrates heterogeneous features such as image, audio, and text representations. In this work, static data are collected at login, temporal logs are streamed during gameplay, and all data are stored in a Hive data warehouse on an HDFS cluster. Batch processing tools like Spark generate high‑density features for offline training of a deep regression model that predicts frame rates.

Application Scenarios

The primary use is to improve game smoothness: when the model forecasts an imminent frame‑rate drop, the engine can lower graphics quality or defer non‑essential resources to keep performance above a threshold. Additionally, the model helps uncover key factors causing drops by interpreting predictions (e.g., via Monte‑Carlo or masking methods), revealing optimization bottlenecks.

Conclusion

By employing a multimodal model, the authors achieve more accurate frame‑rate predictions, enable real‑time smoothing strategies, and provide insights into performance bottlenecks, while transfer learning reduces training cost for new games.