EPTBot on ladder - Page 2
| Forum Index > BW General |
|
jimminy_kriket
Canada5542 Posts
| ||
|
seRapH
United States9822 Posts
On September 19 2026 00:14 RJBTVYOUTUBE wrote: the maphack ruins the entire bot's purpose. Would have been fascinating were it not for the maphack yeah. Still piques my curiosity enough to see what the games look like but as a study it lost a lot of value. | ||
|
iFU.spx
Russian Federation382 Posts
| ||
|
NiteKat
United States149 Posts
| ||
|
PurpleWave
6 Posts
Meanwhile Pluto is publicly available for download, and this bot turned up just a few weeks after Pluto's first preview version was released. But most of all, us folks who make bots have watched hundreds of Pluto games and tens of thousands of other bots playing. It is the community's universal consensus that it's Pluto. Pluto's play is very quirky and unique. Seems very unlikely to see a completely different A/S-tier bot that places buildings chaotically, makes Marines in TvT, has incredible worker combat micro but can't use Dropships, upgrades/researches/builds everything when winning, builds a bunch of defenses before surrendering, etc. | ||
|
TT1
Canada10069 Posts
| ||
|
NiteKat
United States149 Posts
On September 19 2026 03:23 TT1 wrote: whats the point of having a bot if u dont limit its ability (apm etc.)? if players could do operation cwal and show me the money and the bot couldn't then the players would have free wins too, the bot is competing vs humans so it should obv have limitations Normally the bots have fog of war enforced. The ability to command every single unit every single frame (though units don't act on new orders every frame, it depends on the TR/Latency Frames), the lack of need to actually select units, the lack of being limited to the screen's view, not needing to use mouse/keyboard etc., are just design choices from when BWAPI was first designed and written. I wasn't in the scene back then so I cannot speak to why these decisions were made (technical? a personal choice? I cannot say). What I can say is that even if you limit a bot on your own (which you could do, you could simulate a screen and limit information to just that virutual screen, you can enforce unit selection on similar terms to the game client - Pluto enforces a unit selection but I don't think it's screen limited), you could simulate hotkeys etc., APM is STILL not going to be accurate because the API (at least in 1.16.1, no idea what this cheater built to play it in SC:R) still sends commands at the individual unit level, meaning a command to a group of 12 "selected" marines is still going to end up being 12 different actions on one frame instead of, for a human player, 1 command on that frame. It's also a matter of "look at what it takes to mimic a fraction of our power" type of situation - Pluto was a huge investment in the author's time (and presumably money) to get to a level where people will now complain that it should have restrictions put on it to be "fair". Work in the easier domain first then move yourself to the harder domain. Additionally, for the longest time the AIs only ever really played each other. I dunno, just some thoughts. | ||
|
Jealous
10370 Posts
On September 19 2026 03:23 TT1 wrote: whats the point of having a bot if u dont limit its ability (apm etc.)? if players could do operation cwal and show me the money and the bot couldn't then the players would have free wins too, the bot is competing vs humans so it should obv have limitations The point was, and for many authors probably still is, to make the strongest possible AI. Up until recently, as in this specific AI, AI were not able to beat top level players even without limitations. So, why limit an AI if it can't even beat humans? Maybe now that will change, but I guess that is up to the bot author. Think about other AI in other areas. AI has been crushing humans in chess since Deep Blue beat Kasparov like 30 years ago, and they haven't stopped trying to make better AI since. The AI arms race in chess rages on to this day. It doesn't have to have anything to do with human limitations for it to be a goal to pursue. Now there are AIs for chess of all difficulties because some were limited specifically for vs. Human play and training (think Martin on Chess.com). But, the primary goal was always to make the best AI player possible. As a result of advances in chess AI, human players have become stronger as well, by trying to understand the moves the AI is making and practicing against them. To get to that point, it was only possible by pursuing the best possible AI, even after it had crushed all human competition. | ||
|
PurpleWave
6 Posts
Like mentioned earlier Pluto runs at a fixed 240 "APM" which gets counted differently for technical reasons: 12 units being commanded together get counted as 12 actions (therefore overcounting), and BWAPI can not actually select units (therefore undercounting). In addition to being the strongest bot by a wide margin Pluto is the first strong bot with a human-level APM cap. And yes, bots usually play with ordinary vision rules. But the ladder cheater also has enabled a maphack (which actually confuses Pluto; it's unclear whether it plays better or worse as a result). But ultimately the bots have historically been developed primarily for bot-on-bot competitions, which do not enforce APM limits. So that's what they're optimized for (in multiple ways), not playing vs. humans. | ||
|
iFU.spx
Russian Federation382 Posts
The fact that someone used your openly public weights and ported it to remastered during few weeks is actually a very good thing. At least it tells that it CAN be ported to remastered. Please consider me as first in waiting list to play this bot in remastered via your official solution :D I really enjoy watching this bot play. At 12:40 there is a protoss attack and bot sends 19 SCV (!) drilling right into the protoss army. This is a great decision really. https://repmastered.icza.net/game/PB68YIfuDj8pkf0mxbHB-dWqLPRZbiuyQHJo50_AQLs?tab=commands Check frame 18167, there was 19 scv selects and 19 right clicks to mineral patches (those patches are not possible to click, coz no vision, map hack confirmed), but the fact that it made this decision... really cool! | ||
|
iFU.spx
Russian Federation382 Posts
And please add a separate block to the paper with analysis that confirms that when this bot plays protoss it uses only one neuron :D | ||
|
NiteKat
United States149 Posts
On September 19 2026 04:06 iFU.spx wrote: Huge congrats on RL in starcraft! This is very impressive achievement! The fact that someone used your openly public weights and ported it to remastered during few weeks is actually a very good thing. At least it tells that it CAN be ported to remastered. Please consider me as first in waiting list to play this bot in remastered via your official solution :D I really enjoy watching this bot play. At 12:40 there is a protoss attack and bot sends 19 SCV (!) drilling right into the protoss army. This is a great decision really. https://repmastered.icza.net/game/PB68YIfuDj8pkf0mxbHB-dWqLPRZbiuyQHJo50_AQLs?tab=commands Check frame 18167, there was 19 scv selects and 19 right clicks to mineral patches (those patches are not possible to click, coz no vision, map hack confirmed), but the fact that it made this decision... really cool! Unfortunately, you've already been beaten on the count of first to play it. I have some show matches casted where the author was running it in SC:R in private games against a few people. I have a video I need to put up of a showmatch vs. G5 still, which happened on a different date, but you can watch the show matches that preceded it here: The author has also decided to put a hold on show matches for now due to this situation though.  | ||
|
Puosu
7066 Posts
https://arxiv.org/abs/1811.08568 They seem to be affiliated with Meta / Facebook research. Congratulations to them. Big achievement. Too bad it got abused like this. | ||
|
iFU.spx
Russian Federation382 Posts
I look forward to play it on remastered someday anyways. If you (the guy who trained the RL bot) read this, please know, you are the GOAT, this is like watching AlphaGo documentaries really. | ||
|
iPlaY.NettleS
Australia4450 Posts
| ||
|
Puosu
7066 Posts
On September 19 2026 04:28 iFU.spx wrote: Ah, I watched these games, but didn't recognize it was remastered, was too focused on game play) I look forward to play it on remastered someday anyways. If you (the guy who trained the RL bot) read this, please know, you are the GOAT, this is like watching AlphaGo documentaries really. FBH reviewed a number of the replays and said something like: "This is how Lee Sedol must have felt" On September 19 2026 04:29 iPlaY.NettleS wrote: I'm guessing the bot is using mineral boost all game across all bases? Nope. No boosting and it even misplaces expansion CCs in some of the ladder games. | ||
|
PurpleWave
6 Posts
On September 19 2026 04:29 iPlaY.NettleS wrote: I'm guessing the bot is using mineral boost all game across all bases? Most top bots use a few mining optimizations all game (locking, boosting, and -- in the case of Stardust -- return cargo timing) but I have not noticed Pluto doing them. An easy tell that it's not mineral locking is seeing workers in saturated bases go laterally from one patch to another, like human-controlled miners do. Pluto's self-imposed APM limitations (one "command" per six frames) would in theory prevent or greatly limit all of these methods. | ||
|
TMNT
3347 Posts
| ||
|
Vasoline73
United States8092 Posts
It's amazing the bot is MADE though, I want to see non-maphack games against Korean pros ASAP! | ||
|
Connor56205
33 Posts
On September 18 2026 22:29 bmnielsen wrote: I found Paralyze's stream after I posted this and gave them some info about it. Interestingly, internally Pluto runs on 240 APM (its model does a single action every 6 frames, where an action could be selecting a bunch of units). However, between its model and the game, the actual commands get split off into separate commands for each unit, so the APM looks higher. So as an example, on frame 6 it'll select all its workers, then on frame 12 it'll command them to move towards the minerals, but what is actually sent to the game engine is four move commands (one for each worker). The really sad thing is that the hacker added a maphack, so there's no legitimacy to any of the games it is playing right now yo, could you link me the stream he played this bot please? i can't find it. EDIT: sorry i found it | ||
| ||
