Federated Learning–Based Player Personalization in Real-Time Strategy Games : Player Action Prediction in StarCraft II

Sammanfattning

Background: This thesis investigates next-action prediction for StarCraft II macro-management using the SC2EGSet professional replay dataset, a challenging bench-mark with 387 distinct action labels, severely imbalanced class distributions, and rich temporal structure across player-partitioned replay traces. Because player replay data encode behaviourally sensitive patterns, the study frames federated learning as a privacy-aware alternative to centralized data pooling. Objectives: The thesis compares centralized GRU, LSTM, and Transformer sequence models against federated FedAvg and FedProx variants across three Star-Craft II races, and evaluates a personalized backbone–head split architecture in whicha globally aggregated encoder is combined with player-local output layers. Methods: All metric values are extracted directly from stored run artefacts usingofficial train/validation/test splits and final test summaries from all 30 completed runs. Each professional player is treated as a separate federated client, so that FedAvg and FedProx are evaluated under authentic non-IID partitions rather than synthetic client splits. The backbone–head architecture keeps the sequence encoder on the server while player-specific output layers are trained and retained locally on each client. Evaluation covers direct Top-1 accuracy, macro-F1, balanced accuracy,Cohen’s kappa, and communication cost; each configuration was trained with a single fixed seed, so reported differences are best read as point estimates rather than statistically significant effects. Results: Centralized models achieve the highest exact direct Top-1 accuracy (upto 44.0% on Protoss). The best federated Transformer narrows this gap to approximately 5 pp on Protoss and 4 pp on Zerg, and the Transformer ranks first under federated training on all three races despite offering no systematic advantage under centralized training. FedProx provides negligible improvement over FedAvg at the chosen proximal coefficient (μ = 0.01). Backbone–head models, trained and evaluated jointly across all three races, underperform per-race federated baselines on exact accuracy in the completed all-races runs, but retain a meaningful data-local privacy advantage by keeping player-specific output-layer weights on each client and never transmitting them to the server. Hierarchical exact accuracy is uniformly low (under 2%) across all paradigms due to the cascading coarse→fine bottleneck, while the direct exact-action output consistently recovers strong predictions. Class imbalance is the principal bottleneck throughout, making macro-F1, balanced accuracy, and Cohen’s kappa necessary complements to raw accuracy.