While trying (and failing) to reproduce the conditions of this bug with my algo playing against itself in the playground, I ended up seeing an asymmetric behavior from my deterministic algo. Upon inspection of the debug log, I think I guessed what happened:
I think when several firewalls are destroyed in the same frame, particularly when they are destroyed by an unit self-destruction, the order in which those destructions are written in the game_state_string is asymmetric (ie it differs depending the side). I implement an
on_unit_death() method which help my algo predicting it’s opponent’s defense (enemy will probably rebuild recently destroyed firewalls). So the two sides thought their enemy was gonna rebuild a group of recently destroyed destructors over the following turns, but imagined it was gonna be in a different order (I know that from the debug log), and so adopted a different strategy.
Of course it could also come from a bug in my algo, but I’m pretty confident that it’s not the case.
I don’t know if that ‘asymmetric order of deaths in the same frame’ should be considered as a bug, if it is a bug I don’t think it is too problematic.