First of all, I want to thank Magus, Agade, pb4 and reCurse for the time they put to build an engaging contest, thank you guys !
I ended up #46th.
This was the first time I used a search algo with simulation in a contest, with all my previous attempts relying on heuristics/if spaghetti. I used a GA algorithm.
The only thing I want to add to the discussion is the relevance of the search algorithm. I spent most of the contest twiking the evaluation function score parameters, writing more and more complex and creative ways to score the board. All of this with almost no changes (and very often with very bad results) in my AI performance.
So, what did make a difference?
I was getting really frustrated by not getting past wood 1 after a lot of work in the eval function, and then I changed my GA depth from 6 to 3. Magically skipped 2 leages to reach silver. I couldn’t belive it.
Then changing the pool size from 50 to 6 gave me another boost, and finally on saturday I made another change that allowed me to reach Legend: improve performance by ignoring every collision that don’t involve my own looters.
I still couldn’t believe the difference between the top players score and my own (it will end up being like 11+, which is a lot) so I was eager to read their postmortem. After reading Agade’s (which is always very interesting, thanks @Agade) I think I can conclude that the evaluation function is absolutely secondary (in fact I tried Agade’s eval in my GA, and my version beated it 88% of the times), with the search function optimization being what makes all the difference. It doesn’t matter if you have an optimum that is very high if you can’t ever get close to it, right?
That is really the most important thing I learned in this contest that I will try to put in practice in the next ones.
Thanks again to the creators and the CG team for the amazing experience and learning opportunity !