Survival Sequences: Win Prediction from a Strategy Sequence Approach

An important characteristic of battle royale games like PUBG is that the safe zone shrinks as the game phases progress. This makes a player's phase-by-phase strategy critical to their survival. Existing PUBG win prediction studies rely on post-game statistics, which cannot capture the dynamic environment of the game. To overcome this limitation, we proposed two methods. First, we analyzed player behavior from two perspectives: Fight Element and Active element, with a focus on phase-by-phase adaptation. Second, we propose a deep learning win prediction model that utilizes the strategies derived at each phase. Using massive game log data from PUBG, we conducted a data-driven analysis to derive the relationship between phase strategy and eventual win, which we then used in a win prediction model. We trained three time-series deep learning models LSTM, GRU, and RNN to learn strategy sequences over the phases. All three models performed well in predicting wins, with over 88% accuracy.