Профиль

i'm honestly pretty shocked it got this good too considering what i was working with a few months ago, i expect that i can make it even better with a bit more training too

the model is playing as both characters here

as a comparison, this is what the model was doing a couple months ago bsky.app/profile/topt...

different type of training, it's used to make models that predict stuff better at actually accomplishing a specific task instead of just predicting stuff bsky.app/profile/topt...

for those who dont know what "RL" means: its reinforcement learning. the models i was showing before were trained to mimic player behavior as accurately as possible, this one has a second phase of training which tells it to estimate which decisions were "good" and focus on those actions

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

this is correct! thanks, didn't think to check how RISC decayed, will have to update the instruction name to make it more obvious

things i've tried so far: - getting hit/hitting an opponent with varying RISC values - blocking/making an opponent block with varying RISC values none of them have resulted in any change the the meter gained by either player, i'm quite confused, because it varies by character, i assume they use it

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?

i think the only thing i'm lacking now is RL to give the model the goal of winning, not just pure imitation of player habits. also more data, there is really no limit to the amount of replay data i would like to have, it's clearly still overfocusing on a few weird rounds (e.g. the backdash habit)

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

if he does things like 5K>5K often enough that could definitely impact the APM, it's counting everything here including when the opponent is in hit/blockstun

unsure if it meaningfully impacts chipp more but this metric doesnt capture a state transition from a state to itself, that would invalidate things like multiple jabs. might fix that and see if anything changes

i'd guess it's due to low sample size, i think if i ever get more replay folders from the less common characters it'll tighten up to a better average

no actually! thats why i'm so surprised by it bsky.app/profile/topt...

zato and jack-o players please send me your replay folders

in this case i actually expected zato to be unusually low, as eddies attacks wouldn't be counted because theyre a separate entity changing state, not zato. there must be some other thing boosting him unusually high, that's what happens when you have few samples of some characters

this isn't a perfect metric, it over-focuses on attacks, probably has some noise because a button press doesnt necessarily cause any given state change, and the average probably gets raised by stuff like mashing FD, i'm surprised at how clean it seems to be for characters with a lot of match samples

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

still only at ~30 replay folders. if anyone has time to send some over i would greatly appreciate it!