Winner prediction for real-time strategy games through feature selection based on a genetic wrapper

We propose a method that can predict the game winner using a long short-term memory (LSTM) model that is trained using the match data of a real-time strategy game (Clash Royale). Feature selection based on a genetic wrapper is utilized to identify an excellent feature subset. Subsequently, the performance of a model trained using the data of the aforementioned subset is compared with that of an existing model. The model, which is trained using data that includes the entire features before feature selection, exhibited a winner prediction accuracy of about 60%. Conversely, the model, which is trained using the optimal subset data identified through the application of the genetic wrapper feature selection, exhibited a winner prediction accuracy of about 75%. Based on this comparison result, it was found that the proposed method increased the winner prediction accuracy by about 15% compared to the existing method. In excellent chromosomes explored through genetic wrapper feature selection, the spatial feature was always selected. Simultaneously, the temporal feature was always excluded.