I have notice that ranking top of The Great Escape is very static. I was 4th since 9 Feb with pretty ordinary solution and was waiting for my shifting to the down. It looks very strange for me after PR2, which required continuous improvement. Yesterday evening, after move on the 5th place I had decided to submit my new solution - 18th place. After resubmit of my previous version I was on 26. It was exactly the same version which placed me on the 4th. Actually my code, as shown by tests, significantly worse codes from leaders, but how do I was in the top-5 of almost 2 days?
I think this happens because once your bot reaches top 10, the amount of games it plays with different players is much less than bots from places 20 and lower.
Had a pretty similar situation. “Old” TGE bot - 12th place, improved version - ~35th, old resubmitted - ~80th.
Improved version ranked worse ^^
We have to find out how this system works
When a player submits his code to the arena, we select players all over the leaderboard (with jittered sampling) to have a first idea of the ranking. After this we only select players around the actual ranking (± 20) until a player has done 100 games. To be fair, for each game we do all possible combinations. To update the score we use the TrueSkill ranking algorithm.
As we go along, the certainty of a player ranking increases. That’s why it’s harder to decrease in the ranking when you are in the top 10 for a long time. The only games you play are with newcomers in the top 20. We should do some games with players with a much lower ranking . We are actually thinking about this kind of workaround.
Thank you for your feedback.
So can you explain this ?
In my last battles history, my games from 4421 to 4450 are against oyefremov(rank 3) and ecc(307).
Obviously, we have same battles for oyefremov from 4362 to 4391.
A new comer against two top 10… and 29 games in a row.
If after all games, a good rank give no points and a bad rank reduce score, it’s not a problem.
If not, it’s very unfair.
Even with low score you can meet somebody from top-10 immediately after their new version submit.
Once during PR2 I got very bad score after testing contemporaneously with another top 10 player - our AI repeatedly confronted
Do I read correctly that when your 100 “ranking” games are done, you only play against people if they enter TOP20 ?
So if nobody submits a new bot capable of hitting TOP20, your ranking will stay stuck ?
Shouldn’t there be “regular” ranking games. For example, every hour every player would have to replay 5 ranking games with people close to their rank ?
Good idea! I think this will solve the current ranking problems.
I have a different question about multiplayer battles. For 3 players we play 6 battles. Do we get ranked for each battle? Or does the system sum up your scores for those 6 battles and perform a single ranking based on that sum? I prefer the second option, because we don’t want 3-player battles to outweigh the two player battles.
Also, what is the percentage of 2-player battles?
Web introduced this system during the night. For the top 50, a new battle is started every 30 minutes (5 battles for the top 20)
It’s the second option. We play the 6 battles, we get the global winner of this group of battles, and we give this final result to the ranking algorithm (trueskill)
There is around 50% of battle with 2 players, 50% with 3
I wonder if the ranking is frozen immediatly after contest is ended ? Or new solutions have some amount of time to reach appropriate place ?