Question Is it normal ELO Drop down and Up based on Agent swap steps?

I'm using DodgeGame demo. the orignial one , ELO always increasing during self play.

When I have different agents setup, what happen is:

Each swap steps , The rewards are flipped, like GroupMeanReward become negative and ELO decrease, then after 500K steps , the GroupMeanReward become positive and ELO increase.... then after another 500K, GroupMeanReward become negative and ELO decrease.

Is that normal? why it does not act the same?

 

Up

 

Anyone can help?

 

sounds similar to https://forum.unity.com/threads/self-play-oscilates.1380855/ ?

 

The post owner talked to me, and told me he does not remember exactly the reason and mentioned those steps.

I'm trying to carefully checking my code, but I dont use any find function on any level

 

Is there anyone from unity staff can help here?

 

Hey, Did you happen to find any solution yet. I am making a 4 player turn based card game team of 2 vs 2. While training what is happening is my ElO increases with +ve mean group reward but at 200000 step, when there is team swap, it starts decreasing rapidly with -ve mean group reward. It flips every 200000 steps. Please help me if u find any solution to it.

 

Check the players Tag, And Tag order in the observers... I Believe my case was Both team accidently has the same tag and same observation order

So Blue Team should have "blue" tag , and the sensor that see the enemy should see the enemy team first like in tag order "purple , blue"

And the other team should be "purple" and observation order should be "blue" then "purple"

I think that was our mistake... you can revert to original prefab from dodgeball example to verify your setup

 

Thanks for the reply, I did checked it many a times. Since mine is a turn based card game, Observations are like :-

My position on table
fist turn
My points so far
opposition points so far
My cards (with cards details and points of each card)
Cards on table (with card details and points of each card)
Total payed card(history cards)

So, its same for every team, Blue team gets their points first and then purple teams, and vice versa.
Dont know still why would the elo drop.

Also to add, both team needs to make different points to win, like if total points is 32, Blue needs to collect 20 to win and Purple needs 13 to win, whoever makes that fast, wins.