One of the approaches applied by a number of teams participating in the Google Research Football competition on Kaggle was Imitation Learning. Over time as the competition progressed, there were more and more game replays available to train on. If you want to pursue this research direction and need data to start, Kaggle competition replays are there for you. It is possible to collect the replays of your interest using Kaggle API or you can start with an already pre-processed publicly available data set. An example of such a dataset can be found here. It consists of around 5000 replays mainly from the top competition agents. Replays were converted from the original Kaggle format to reduce the size of the dataset.