Multi-Task Learning for Heterogeneous Prediction from Video Game State with Transfer Learning

Research Question and Approach

This study investigates whether multi-task learning (MTL) can effectively consolidate multiple prediction tasks—such as win probability, damage estimation, and movement prediction—into a single, shared model for team-based multiplayer games. The authors aim to determine if joint training improves generalization and reduces computational costs compared to maintaining separate models for each task. They utilize a large-scale proprietary dataset from World of Tanks, employing a multimodal architecture that integrates rasterized map images, global match context, and per-unit telemetry through an attention-based encoder.

Methodology

The researchers implement a hard-parameter-sharing architecture where a shared encoder extracts features that are then passed to task-specific heads. Because the tasks involve mixed modalities (binary classification and regression) and varying loss scales, the study evaluates several optimization strategies to balance task influence. These include equal weighting, random loss weighting (RLW), and gradient-balancing methods like PCGrad and FAMO, which are designed to resolve gradient conflicts where optimizing one task might inadvertently degrade performance on another.

The study demonstrates that MTL is a viable and efficient approach for game-state prediction. By training a shared representation, the model achieves better performance than isolated single-task models. The authors find that naive loss summation is often insufficient due to scale mismatches and conflicting gradients between tasks. Advanced gradient-balancing techniques, such as PCGrad, prove effective in navigating these conflicts, leading to more stable training and improved overall accuracy. Furthermore, the shared representations learned through MTL show strong potential for transfer learning, maintaining utility even when applied across different game maps.

In modern game analytics and AI, telemetry data is abundant but often siloed into specialized models. This research provides a practical framework for developers to consolidate these models, reducing inference latency and training overhead without sacrificing predictive performance. By demonstrating that MTL can handle heterogeneous tasks effectively, the paper offers a blueprint for building more efficient and scalable game-intelligence systems.