BazkieBumpercar
commented
4 years ago Thank you for your question ziaoang!
My defensive system consists of various methods in a somewhat overly complicated combination. It was a while ago when I programmed that so I can't really explain it all too well. I think it's a combination of:
- hungarian algorithm, to find the best 'fit' for the current player's positions, into the formation setup (who is most fit to be left-back right now? or keeper? etc)
- man-marking system, where the most dangerous opponents are assigned to the player that have the most chance of reaching them before they have a chance to shoot
- magnets/anti-magnets system so that people are attracted to the ball, keep proper distance from teammates, etc.
- some custom code to glue this all together in various ways
One thing I do remember is how much harder defensive systems are, as compared to offensive systems. It makes sense; an offensive mistake is not so bad, while a defensive mistake often leads to a goal.
I put a lot of time and effort into coming up with the eventual code, and I too am very happy with the result, it's pretty solid.
I haven't looked at google's code, so I can't say much about it. Maybe the difference is that I have a lot of experience with watching, playing, and thinking about football == experience, while they are programmers first, not football fans (I'm not sure about that, of course :)).
I think that a machine learning AI could potentially be better at football defense than my rule-based code. But the latter is still preferred for games (for me at least), since it's easier to change, maintain etc (since you can control every part of it. neural networks are more 'black box-y')
ziaoang
Hi BazkieBumpercar, your work is introduced in Google Research Football. Google trains a lot of AI to challenge your rule-based AI, but can not defeat your hardest AI. How do you design your rule-based AI, is there some papers or blobs? Thanks~