EPTOn September 22 2026 01:06 [sc1f]eonzerg wrote:
any replay with pluto playing zerg ? :D
any replay with pluto playing zerg ? :D
been looking for that too, all i could find is this, but its not pluto
Bot on ladder - Page 7
| Forum Index > BW General |
|
Crimson)S(hadow
Philippines1083 Posts
been looking for that too, all i could find is this, but its not pluto | ||
|
cha0
Canada518 Posts
On September 22 2026 01:06 [sc1f]eonzerg wrote: any replay with pluto playing zerg ? :D Not against a human but this is from the finals: https://dgant.github.io/openbw-replay-viewer/?rep=https://davechurchill.ca/starcraft/cog/results/2026/replays_final/00004-PLUT_PURP.REP Pluto as zerg vs PurpleWave as toss | ||
|
quaristice
177 Posts
| ||
|
Jealous
10416 Posts
On September 22 2026 02:17 quaristice wrote: occurs to me that the bot can't forget to rebuild a lost floating barracks or ebay or make mistakes like that bc it's always got awareness of its build options That might be a bit of a stretch IMO, but would need confirmation from someone with more knowledge. In my experience up until now, this is precisely the kind of thing that humans are better at than AI. AI doesn't "remember" or "know" things like that on its own without being "told," usually? Maybe this one is different but historically it has felt like the case. So, in theory possible, but not a guarantee. | ||
|
quaristice
177 Posts
On September 22 2026 02:27 Jealous wrote: That might be a bit of a stretch IMO, but would need confirmation from someone with more knowledge. In my experience up until now, this is precisely the kind of thing that humans are better at than AI. AI doesn't "remember" or "know" things like that on its own without being "told," usually? Maybe this one is different but historically it has felt like the case. So, in theory possible, but not a guarantee. from what i understand from reading the author's explanations is that when it looks at the possible orders it can give "build a factory/turret" is either enabled or disabled and it's always looking at all its possible orders it can give. it's part of the state of the world that it's "told", to use your wording | ||
|
PurpleWave
24 Posts
Rule-based bots can't really "forget" anything because they are not limited by attention. PurpleWave, for example, knows everything it wants to make for the next few minutes (though the preferences may change) and has pretty good estimates on what frame each of them will be able to start. If having an ebay is considered more important than the next tank, PurpleWave will replace the ebay immediately and know exactly how delayed the tank will be as a result (which may then impact when it needs to add the next supply depot, etc). The whole economy is simulated minutes into the future. @cha0 Lots of replays here of Pluto as all races against other bots from the COG tournament it just won. | ||
|
Jealous
10416 Posts
On September 22 2026 02:41 PurpleWave wrote: @Jealous Any good bot in the last decade would replace a lost building if it needed to. Not a category of error you see. Rule-based bots can't really "forget" anything because they are not limited by attention. PurpleWave, for example, knows everything it wants to make for the next few minutes (though the preferences may change) and has pretty good estimates on what frame each of them will be able to start. If having an ebay is considered more important than the next tank, PurpleWave will replace the ebay immediately and know exactly how delayed the tank will be as a result (which may then impact when it needs to add the next supply depot, etc). The whole economy is simulated minutes into the future. @cha0 Lots of replays here of Pluto as all races against other bots from the COG tournament it just won. Cool, always glad to be corrected on stuff like this 🙂 I guess "good" is the operative word here; working my way up through SCHNAIL a couple years ago, I recall seeing instances where the AI would be floating resources instead of replacing lost buildings, so I took the broader view. Furthermore, given that Pluto came from a different development background than other AI, I didn't want to jump to conclusions, though it seems I ended up doing just that. Thanks for clarifying, and apologies to quaristice for my error 🙏 | ||
|
CostaCustoms
2 Posts
On September 21 2026 03:29 PurpleWave wrote: That's what SCHNAIL is. It would be reasonable to offer it (consensually) on either Battle.net or Shield Battery if: - There existed a Remastered client with bot support but *without* enabling ladder abuse (this is tricky) - Someone is willing to foot the bill for servers which host and run the bots Bots generally require a dedicated CPU core per game for adequate performance. Pluto requires 8 CPU cores or a GPU (does not have to be modern). That's on top of needing to host the remote instance of StarCraft. These constraints explain the design of SCHNAIL: It's in 1.16.1 and the bots run locally on your machine. That has its own downsides (antiviruses/Windows Defender frequently quarantine BWAPI or its bots. And some players don't have adequate hardware to run the bots) | ||
|
TelecoM
United States10893 Posts
| ||
|
Connor56205
58 Posts
| ||
|
RowdierBob
Australia13537 Posts
| ||
|
AttackZerg
United States7700 Posts
(The worst thing about the Pluto revelation is now PurpleWave bot is gone from schnail!) From what I've seen, Pluto is still within reach of and potentially has larger more targetable gaps then other bots. Eventually we will see "flash finetunes" "Jaedong finetunes" of models. | ||
|
TelecoM
United States10893 Posts
| ||
|
NoS-Craig
Australia3173 Posts
| ||
|
iCCup.Trent
Argentina469 Posts
| ||
KwarK
United States44439 Posts
On September 22 2026 08:52 Connor56205 wrote: I posted Paralyze fpv of all 7 games vs Pluto https://youtu.be/6OwQw0ePrB8 Thank you for posting this, this is really interesting. | ||
|
Kraekkling
769 Posts
are there more details on how Pluto was trained (gpu hours) and/or is there a repo for the training? | ||
|
KwarK
United States44439 Posts
On September 22 2026 10:38 RowdierBob wrote: Its micro is insane. Paralyze had the perfect army comp to take out take out the bio comp but its micro is just way too good. God forbid it learns how to macro properly and play mech vs P. Or maybe bio is the ideal comp vs P and humans just aren't good enough to control it properly. Individual tanks were retargeted onto reavers each time the reaver unloaded. Same with the ht comps, the tanks just said "nope" and wiped out every spellcaster in one volley with the right amount of tanks targeted onto each. With mech the vultures would probably only target zealots while the tanks stayed on dragoons. Zero drop defence seems to have been a recurring weakness though. The loss after the game state at 1:08 in the VOD is insane but the high templar just couldn’t get storms off. | ||
ImbaTosS
United Kingdom1749 Posts
| ||
|
quaristice
177 Posts
| ||
| ||
