Learning Action Evaluation Function for Soccer with Human-Creatable Training Data

Hidehisa Akiyama

In competitive team sports, individual player decisions significantly impact
overall team performance. Designing an appropriate evaluation function for scoring a
player’s action selection in complex games such as soccer is a challenging problem. It
is essential to reflect the supervisor’s instructions in decision making, but if a human is
assumed to be the supervisor, it is difficult for the supervisor to accurately score a large
number of trials. This paper employs a learning to rank method to obtain an evaluation
function for action selection, focusing on ball-chasing behavior in soccer. We employed
a gradient boosting decision tree as a learning to rank model. The RoboCup Soccer Sim-
ulator is used for experiments, deriving a ranking model from players’ action logs. The
results show a model with satisfactory performance can be learned when the number of
situations exceeds approximately 1,000, even with training data that is assumed to be
created by humans.

 ICIC Express Letters Part B: Applications

15

6

593

600

10.24507/icicelb.15.06.593