Fig. 3 compares the accuracy of the strategy prediction over
game times. In the results, the “fog-of-war” is considered. If
scouting is not available to the players, the amount of
information about the opponent’s behavior should be very
small and makes difficult to predict their strategy. The results
show that the human’s replay logs are more predictable than
the one from the Bots. The gap is bigger in the early and later
stages of the games than the middle stage. Fig. 4 shows the
utility of machine learning prediction. If the player changes its
build order based on the prediction, the estimated win ratio is
calculated. It shows that the human’s are not successful to get
benefit from the change of build orders based on the
prediction. It means that humans already consider the
relationships among the build orders. However, there is still
enough room for the improvement by applying the build order
change based on predictions.