Pangaea

@toptwentynotes.bsky.social

hi, i made those ArcSys script modding tools, and the Xrd SAMMI/WebSockets integration. posting stuff i make here. @TopTwentyNotes on twitter Support my work: https://ko-fi.com/pangaea__

Xrd model now uses RL, this isn't a finished training run, and no online RL has been applied yet as i'm still working on the mod setup for that, but i'm very happy with the progress! here's the model playing against itself, i think it's a massive step up from the last footage i posted

working on figuring out the meter gain values for each character, i've created a table of each characters values, most of them are fairly obvious, just meter added and scaled by tension pulse somehow, but i can't figure out what "Guard Balance Tension" is supposed to be. any ideas?

A table of meter gain multipliers by character in Guilty Gear Xrd

switched to training on whole rounds instead of small context windows, i'm pretty shocked by how big the improvement is compared to what i had before. this also enables performance optimizations which make it way easier to run larger models in realtime. here's a Johnny vs Faust round as an example

trying to figure out how to measure the average APM of each character in the Xrd dataset. here i'm searching for changes in the player state that coincide with a button press. honestly not sure what's going on with zato

A plot showing the average button press-correlated state changes per minute for each Xrd character, this is supposed to roughly indicate the average APM of each character in the dataset

if anyone has more Xrd replay data folders they're able to send over, i would appreciate it, going to try and continue work on the data science stuff i'm doing, but i want to get the most representative sample of Xrd play that i can

the input prediction model takes the state that characters are in as input. these are placed in a vector space, starting out in random positions, but are slowly positioned so that states that are "related" end up closer together. here's a visualization of the random placement vs the sorted result

All states contained in the Xrd dataset, placed randomly throughout a vector space and projected to 2D with UMAPAll states contained in the Xrd dataset, placed throughout a vector space so that related states are closer together and projected to 2D with UMAP

new BBScript version, this adds support for the latest Guilty Gear Strive update. it also stabilizes some changes kkots made, which fix formatting for scripts where ArcSys devs didn't close scopes correctly, as well as adding much more Xrd info github.com/super-contin...

Release v1.5.8: Support latest GGST version, stabilize other changes · super-continent/bbscript

This update adds support for the latest version of GGST, as well as stabilizing changes from 1.5.7

github.com

if anyone has more Xrd replay data folders they're able to send over, i would appreciate it, going to try and continue work on the data science stuff i'm doing, but i want to get the most representative sample of Xrd play that i can

new model, this time it's trained to predict further ahead in the future, hoping that if i tweak this right it will learn to plan ahead better and maybe execute real combos, there's already some improvement to sols short strings but it doesn't seem to be enough yet

early version of a new model, seems to be a bit better at short combo strings but still cant do much more than that, also still has some pretty silly decision making in neutral. its starting to feel less like a bot and more like a very weird human player which is cool though

after trying to debug the 5D spam assuming it was an issue with the model itself, i've realized it's because i forgot to change the Xrd controller binds back to the default after playing a bit, meaning the program automating the inputs was pressing the incorrect buttons the whole time

Pangaea@toptwentynotes.bsky.social · 4mo ago

been working on player input prediction using the replay dataset. both players here are bots. still a rough result, but i've made a lot of progress since starting

if you would like to help build this dataset, i would greatly appreciate anyone sending me more replay folders, especially the ones for any character with <100 appearances! just compress the "REPLAY" folder inside the "SAVE" folder in your Xrd install and send it over

The total number of times a character appears in the current Xrd dataset:
Faust: 793
Ky: 642
Raven: 569
Sol: 448
Slayer: 388
Dizzy: 351
Sin: 294
Baiken: 257
Haehyun: 230
Elphelt: 209
Millia: 194
Bedman: 113
Leo: 109
Ino: 106
Chipp: 96
Johnny: 82
May: 68
Answer: 54
Venom: 45
Potemkin: 44
JackO: 35
Jam: 33
Axl: 27
Ramlethal: 19

been working on more ways to visualize the xrd replay dataset. here is a density plot illustrating the distance between the character and the opponent when the character uses a grounded normal attack

A chart of the distance between characters during each successful midscreen grounded normal attack. Represented as a density plot

trying to come up with more ways to plot the data i have, here's one match with x/y over time, creates a cool little structure, nothing too surprising here though, i think i should focus on resources and state-related stuff next

Bild

been spending time chipping away at a fighting game framework. currently have motion inputs and states worked out, they are very flexible and take after the concepts i've learned modding arcsys games. no test footage worth sharing yet, but i find the example code satisfying

Lua code describing a forward walk state for Ryu in a test fighting game engine.Lua code defining motion input parsing for 236, a button being pressed, and the composition of the two, declaratively constructing a 236A motion input parser

i'm now up to 10 replay folders total, assuming all are full (Xrd has a 100 replay limit) and use best 2 of 3, that gets me around 2300 rounds of data! i'm always open to more, but this seems like a good start for some testing. thank you to all who submitted their folders!

i'm looking to get as much Xrd replay data as i possibly can. would anyone be willing to send over their folders? i have a mod that can automatically play back the replays and save information about them , so all i need is an archive of the REPLAY folder in the path `Guilty Gear Xrd/SAVE/REPLAY`

Xrd Replay Takeover and Save States v0.2 (link in reply) - Fully integrated into in-game UI/menu/controls - Fast forward and rewind, no longer restricted to the most recent 500 frames like in previous versions. - Fixes for many common crashes

A random thing I found interesting is that every pause menu in Xrd is the same menu. There's a big table that indicates which buttons/sub-menus should be enabled in each game mode. You can edit this table to see them all which reveals some unlocalised and empty menus presumably from development.