Predikce výsledků kola ve hře Counter-Strike 2 za pomoci strojového učení

This thesis develops machine-learning models for estimating in-round win probability in Counter-Strike 2, a competitive first-person shooter with a professional esports scene. The dataset consists of 226,466 game-state snapshots extracted from 764 parsed map demos of nine professional tournaments played across eight maps in 2025 and 2026. A hybrid sampling strategy triggers observations at kills, bomb plants, and round boundaries, supplemented by fixed-interval snapshots during quieter periods. Each snapshot aggregates player-level data into team-level features covering survival state, economy, loadout composition, spatial positioning, and short-term momentum. Ten feature-set variants of increasing complexity are evaluated through a two-phase model-selection procedure applied independently to two model families, gradient-boosted trees (XGBoost) and feed-forward neural networks. Both families select the same 43-feature Spatial set. On the held-out test set, XGBoost achieves a log loss of 0.4272 and an accuracy of 77.85 %, while the neural network achieves a log loss of 0.4294 and an accuracy of 77.76 %. The near-identical performance indicates that both architectures extract the available predictive signal from team-level tabular features to a similar degree. SHAP-based interpretation shows that both models rely most on player survival counts, health totals, and equipment value, with eight of their ten most important features in common. Prediction difficulty decreases as rounds progress, from a log loss around 0.53 at round start to below 0.15 in the final 20 seconds of the bomb timer.