Elo score boom

[pier4r]

flow players vs ghost.

2018 08 25 - 2018 08 31

Code: Select all

sqlite> .read net_flow_players_ghost_points.sqlite
-13637
-11752


so 25389 net points from ghosts to players

[pier4r]

flow players vs ghost.

2018 09 01 - 2018 09 07

Code: Select all

sqlite> .read net_flow_players_ghost_points.sqlite
-3070
-4106


so 7176 net points from ghosts to players

[DjSapsan]

I playing AoE2 with ELO system. My notifications: the more players the less issue. But I can point on important things.
1) Gladiabots is a full determined game. Even random totally determined. Its mean that your bots will fight always identical matches.
2) RTS strategies is a randomized. Topplayer can get tired and lose to weak player and so on. Every match is different.
In general there no difference between matches in Gladiabots and RTS. Players always improving and making mistakes. So their ELO sometimes shaking more or less.

BUT GLADIABOTS IS TOTALLY DETERMINED.
Considering, one player create super strong AI that winning everyone (or 99% at least). This AI will never be tired or sleepy. After few thousand games this AI will have 10000 ELO and other players will drop closer to 0 ELO.
Fortunately, in ranked games we have limits for playing with different leagues. Still, top leagues will lose the players.
Assume, superstrong AIs will destroy entire ELO balance (lowrank still in safety tho)

[pier4r]

"After few thousand games this AI will have 10000 ELO and other players will drop closer to 0 ELO. "

No, it doesn't work like this. Impressive how many people play and see games based on elo and variants and how few understand it. An unbeatable oppoenent surely extend his gap but not thousands points large.

Also the elo score boom is about how many points the active players hold vs how many points they had when they started.

[ElvenMonk]

Inflation has a bit of problem though. I totally agree with Pier's definition of advantage, which is always a relative thing. But leagues are at fixed ELO points, so higher percent of active players will get to higher leagues over time.

If disabling dropping ghosts are the main source of slow inflation, we can introduce a victory tax.

Lets say in GM real people have K-factor = 20 if they lose and K-factor 18 if they win.
It will constantly take some points from the system.
We can play around exact tax rate based on ghost released points statistics.

[pier4r]

For the leagues based on fixed scores it is true. Although with resets and adjustment you can fix it.

In discord people had the idea to take the scores of the previous seasons to set them for the new seasons.

Example in season 1 the 25th place had, say, 1980 points. Then in the new season 1980, or the score of the 25th player (active or not) is the minimum to be a GM.
If seasons are short enough then this method should be quite ok.

Elvenmonky you also miss some understanding how the score formula works. If you do K-factor 20 if one loses and k-factor 18 if one wins, you may have less fast progression towards the top, but the points (defined by the kfactor) are simply exchanged. They are not removed from the game. Also playing with taxes without testing the taxes properly, normally creates more probelms that it resolves. The elo formula is tested over years, a random tax just easily written in English is not.

Then once again, for the simulation I did, the problem of "inflated scores" happens even without ghosts. I let 10 entities play each other for 1800 matches and the strongest one ended up with 4000+ score on an average of 50 runs (so 1800 games x 50 ). (the base score was 1000)
Accumulating score at the top is something that happens, unless everyone has the same skill (and in that case the game would be boring). You cannot avoid it without making the score useless.

I agree on the leagues thresholds, but then I think that using the previous season threshold is a simple change that should work and if it doesn't work is not going to create havoc.

Of course this is true if the seasons happens regularly every 1-2 months (the seasons themselves are quite a tax as they reset the score to the base league score). If not, the idea won't do much.

For example pre reset the top 20 (active or not) had 48355 points. (n1)
Now the top20 (active or not) has: 48876 (this 1 month and 2 days after the reset)
A super close value, so the scores are relatively stable unless one let the system run indefinitely with ghosts + many active players (without active players the points would be frozen).

Also a player has to be stronger than the ghosts to accumulate points (and release ghosts with more points). It is not that trivial.

With this I mean that recurring resets + adjusted league thresholds would fix the perceived problem of "GM league is getting easier". Although the idea that I read here and in discord "ah everyone is a GM now" it is quite a subjective argument. I wouldn't mind having only one league (but with small kfactor, so like 20). Bragging about the league is bragging on epenises, shows little maturity.

n1: viewtopic.php?f=7&t=1354#p16834 (historical data refutes so much easy sentences)

[ElvenMonk]

Example in season 1 the 25th place had, say, 1980 points. Then in the new season 1980, or the score of the 25th player (active or not) is the minimum to be a GM.

Number of people in each league should be relative to total number of active players. I believe it was true for your simulations. Having constant number like 25 players in GM is not fare.

Elvenmonky you also miss some understanding how the score formula works. If you do K-factor 20 if one loses and k-factor 18 if one wins, you may have less fast progression towards the top, but the points (defined by the kfactor) are simply exchanged. They are not removed from the game.

Well, they are, because different kfactors applied to parties of the same match.
Say, if I win player with same score in GM, I get 9 points and he/she loses 10.

Some elo points leak already happens when people play against ghosts from different league. As party from higher league is expected to win, it gets less points, and party from lower league loses more points. Sometimes result can be opposite, but in this case winning player from lower league has all chances to be promoted quickly. So the fenomena of disappearing points on the edge of leagues must exist.

[pier4r]

the phenomena of points migrating from ghosts to players happens at every range. The more active the players, the more it happens on average.

"Number of people in each league should be relative to total number of active players. I believe it was true for your simulations. Having constant number like 25 players in GM is not fare."
It is not what I mean, maybe I didn't explain myself.

The idea of the new league rating doesn't come from me, but from a discussion in discord (that now is very hard to find because chats).

One changes only the league thresholds of master and GM.

If the season 1 ends, say, today, an idea to make GM a bit more valuable is to leave all the leagues as they are except master and GM.

For GM threshold one can pick the actual value of the 25th player (or 20, or 30, or whatever), looking also at the inactive players.
The actual 25th player is:
25 Oiide 2071 (max 2307)

So it means that season 2 will have:
master from 1550 to 2070
GM from 2070 and above.

Actually the threshold is:
master from 1550 to 1800
GM 1800 and above.

This doesn't mean that GM will have only 25 players (or 20, or 30, or whatever). Rather it means that the players to reach GM will need to produce better solutions to get higher scores.

Only those above 2070 will be GM.

Now back to inflation. If left unchecked, over time it will be easier for players to gain 10-20 score points on average as the ghosts get farmed, and therefore even 2070 will be pointless.
But all the discussion about new leagues thresholds presumes that season resets will happen relatively frequently. Actually one should happen this month or at the end of the month.

Then say at the end of season 2 the 25th player (counting active and inactive) has like 2030 points.
Therefore for season3 the threshold will be:
master from 1550 to 2030
GM from 2030 and above.


Same story with season 4. At the end of season 3 the 25th player has a score of 2107.
Therefore for season 4 the threshold will be
master 1550-2100
Gm 2100 and above

and so on.

[ElvenMonk]

Which means you still want to have only 25 players in GM by the end of next season, if they perform on approximately same level as before.

And currently seasonal resets make most GM players to feed underscored ghosts of top10 players first.

I prefer leagues at specific points. Like 1700 for master, 2000 for GM, 2500 for (?)Legend..

[pier4r]

"I prefer leagues at specific points."

Indeed. You just change the point according to previous performance.
Grandmaster until summer 2017 didn't exist. It was created based on the performances in master.

You can also add legend, but the threshold should be based on the previous performances. If you set, say, 6586947000 (made up number) as threshold, it will be always empty. So you pick a value that is likely to be reached.

[ElvenMonk]

If thresholds are based on numbers relative to total number of players for previous season, I'm ok with that.

[pier4r]

For those who wants to get the "maximum score reached by a ghost listed in the stats page".

Copy the ghost matches of a player from the stats page in a txt file. Then run those commands (possibly under cygwin / linux . Don't forget the line endings of the file).

Code: Select all

awk '{print ($(NF-1) + $NF) " " $1}' data.txt | sort -rn -k2 | uniq -f 1 | sort -n