Only create the MR if the model trained on both mcts- and stacking-generated data is significantly better than the previous model
The best thing to do would be to learn on both kinds of generated games. However, it would lead to a cycle: every time the implementation is updated, the generated stacking games would change, so learning on them would change the implementation.
The way out is simple: train the model on both kinds of games. Once the model is trained, build and install a modified version of tarot in /opt/candidate, for instance, and compare the original version with the /opt/candidate version (beware though, we need to use two different libtarot in two different LD_LIBRARY_PATH). If the candidate is significantly better than the original implementation, then submit the merge request. Otherwise, don't do anything.