Fig 1.
3D-ResNet model with non-local neural blocks.
Fig 2.
Architecture of the 3D-ResNet with non-local neural blocks.
Fig 3.
The Grad-CAM procedure in the extraction of a key situation at a certain time point in StarCraft Ⅱ.
Fig 4.
Parsed StarCraft Ⅱ features using the SC2LE.
Table 1.
The confusion matrix of predicted the game results.
Table 2.
The winning prediction model’s performance concerning game time length.
Table 3.
The winning prediction model’s performance with different f functions in non-local blocks.
Table 4.
The winning prediction model’s performance with different locations of each non-local block.
Table 5.
The winning prediction model’s performance with different machine learning and deep learning models.
Fig 5.
Win-loss prediction model performance with different number of sampling frames.
Fig 6.
The ROC curve in our proposed model.
Fig 7.
Protoss win probability per frame using a sliding window.
Fig 8.
Example of the Protoss winning games: Grad-CAM focused on the scene in which Protoss defeated Terran in combat.
Fig 9.
Example of the Terran winning games: Grad-CAM focused on the scene in which Terran defeated Protoss in combat.
Fig 10.
The process of Grad-CAM evaluation based on the highest Grad-CAM score.
Table 6.
Comparison of model performance based on the area corresponding to the Grad-CAM. score.