-
Posts
17.008 -
Joined
-
Last visited
-
Days Won
524
Posts posted by Stan`
-
-
14 minutes ago, Feldfeld said:
Now the graph indicates I'm dropping rating for each match, not sure that's accurate
I'm not sure but I think it's normal due to the very low amount of matches. At the beginning you have a deviation of 350 which probably boosts your rating at each each match, now your deviation is closer to 45. But at least the dates are in order now
Another thing is that there aren't many matches.
The code responsible for this is there https://github.com/StanleySweet/replay-pallas-api/blob/main/src/instant-glicko-2/Glicko2Manager.ts#L64
- 1
-
1 hour ago, Feldfeld said:
Is there a duplicate detection for replays? Ideally it would keep also the longest version available (there could be a somewhat frequent situation where a player spectated a 1v1 then left in the middle of the game, and posted the replay first)
Only the first one remains. The replay matchid is used to avoid duplicate replays.
1 hour ago, Feldfeld said:The glicko data looks inconsistent. It tells me my rating is 1545 while the graph puts me above 1600
Ha I messed up the sorting should be fixed now.
-
Pushed some updates, you now have the glicko rating on the player profile. Also added the game rating info I have to that graph. I didn't exclude 1v1 from local ratings as for now those are two different things.
People with the contributor badge can upload replays.
I haven't set up the scrapper yet, not sure if there I should create a specific thread for it and whether it's a good idea.
- 3
-
IIRC @vladislavbelov worked on it some time ago.
- 1
-
8 hours ago, Feldfeld said:
Yeah that's a bigger issue. A mitigation I see to avoid weird looking swings would be to recompute the whole thing at even slower intervals at which replays are scrapped. So there could be replays displayed not yet used in the rating calculation. So the problem would be "hidden" behind the big quantity of replays at each update. However, it happens that the multiplayer community plays much more TGs than they do 1v1s so not sure it would work well.
Found this little thing that might solve the issue (it's also what lichess uses), but I need to convert the rust code to TS https://github.com/gpluscb/instant-glicko-2/tree/f34fd27aba57c21b0554a6ccc2b6a26559a0f5e7
- 1
-
Regarding determining if a player has lost or not, would the percentage Building Lost + Units lost work? Lowest Percentage wins? (This would be done only for 1v1 and if there is no state set to won)
-
3 minutes ago, Feldfeld said:
Yeah that's a bigger issue. A mitigation I see to avoid weird looking swings would be to recompute the whole thing at even slower intervals at which replays are scrapped. So there could be replays displayed not yet used in the rating calculation. So the problem would be "hidden" behind the big quantity of replays at each update. However, it happens that the multiplayer community plays much more TGs than they do 1v1s so not sure it would work well.
That could work I suppose. I can also add a big button for admins to recompute it. I'll try to implement it. Should we unlike lichess start at 1200 ?
-
Another issue we have is that unlike on lichess, one can upload replays in any order, so I suppose I should recompute the whole thing each time someone adds a new replay for all players (since a changing rating for someone would change everybody that ever played against them)
-
48 minutes ago, Feldfeld said:
No separation
Okay. Do we count replays with no winners as draws ?
https://github.com/mmai/glicko2js
They say there that
You should not update the ranking after each match. The typical use of glicko is to calculate the ratings after each tournament (ie collection of matches in a period of time). A player rating will evolve after a tournament has finished, but not during the tournament.
So I'm not sure it would work well in our case
-
25 minutes ago, Feldfeld said:
My idea would be for 1v1 only both ranked and unranked
Separate ratings for both? Or one ?
Spoilerteam games would use the Local Ratings.
Should it keep the 1v1 then?
27 minutes ago, Feldfeld said:Would need opinions of others to know if it even is a desired feature.
Sure I'll wait for the input
-
Sounds fine (we can still compare it with the in-game rating)
I don't know the maths though, so I'll need some help with that. Also I need to know which replays would be concerned or how many elos one would have.
There is rated/non rated
Teams locked NvN
AI's
Some matches don't have informations about who won and should probably be excluded.If I can gather a good enough specification, I can make the code.
-
4 hours ago, Feldfeld said:
Nice to see it up!
Seems like the website enforces a password of length 9 despite claiming it should be 8 characters.
On the Local Ratings section, the number of matches column doesn't seem consistent with other data. (also I'd argue that for 1v1 some form of elo system is always going to be better than local ratings)
Indeed, thanks for the report(s)!
I do have the game's elo from the names, would you like to see a graph about that ? Plotted by date something like Game Rating Evolution?
-
Published some updates to the dark mode, a few fixes and a new graph.
- 1
-
For now yeah. I may reconsider in the future but I don't like the idea of having the api left unrestricted
-
-
Did you use \ in your changes ?
-
This video explains it well https://youtu.be/C8YtdC8mxTU?t=976
- 1
-
-
6 hours ago, hyperion said:
Smooth shading interpolates geometry during render, which isn't free and so the game engine doesn't do it as far as I'm aware, at least I never stumbled over code that would indicate otherwise. What you get is basically the same as flat shading in blender.
Of course we have smooth shading Oo. Else barrels would look much worse. Units too.
-
I suppose @wowgetoffyourcellphone is in charge now
- 1
-
Okay few things from looking at the DAE in that zip.
You need two UVs in order to be able to use AO. The fact it doesn't crash for you makes me think that either you got an error message, or that you might not have set the material quality to 10.Second is wasted UV space, you want to have as little black spots as possible, you can probably rearange things so that those curved pieces take less space And yeah you can boost the margin a bit.
- 2
-
Can you show me the generated ao and the second uv map layout. Not sure i'll be able to access blender.
Also if you export the roughness map you need to invert colors. That's what I meant by 1-roughness.
- 2
-
The problem is the existing voices wouldn't be enough to train a model. So until legislation is in vigor, I'd be really careful in using AI in anything else than mods.
- 1
-
1 hour ago, ShadowOfHassen said:
I don't know if trac can do something similar, but GitHub has the ability to autogenerate release notes by looking at what PRs were merged.
Yeah you can actually do that just with git log --format.
But the problem is here is that not everything is relevant, you don't want to see 200 autobuild commits
Nor fixes for art typos. Just the art commit itself. They can sometimes be grouped. That's why we (used to) do it by hand.
- 1
Replay Pallàs — A platform to share your replays
in Game Replays
Posted
Found the guilty line
https://github.com/StanleySweet/replay-pallas-api/commit/e085022fc93e7a6578a82d18c7a3df21970a2d5e
Explains the weird link between deviation and rating.
I hope my friend's @Stockfish's mouse is okay, they are clicking at least twice as much as any one
@borg- has a terrible rating because he lost the three matches on the site (His rating is 900? the '?' being for provisional)