|
On September 20 2026 02:17 PurpleWave wrote:Good to see you here Evan. I don't know if you saw that I got Overmind running on modern BWAPI and playing on BASIL. Some clarifications, some from the Pluto FAQ, some from chatting with the author, and some from watching hundreds of Pluto games. Pluto was trained without map hacks and is designed to run without map hacks. It won the COG tournament easily without them. Only the cheater who is playing with Pluto is using a map hack. It does not explicitly encode high level strategies but does have a Thompson-sampled strategy bandit. It was trained with rewards for following the selected build order (but is not forced to do so). The model includes recurrent state to assist with long-term memory. Pluto has generalized impressively well against out-of-distribution play. For the COG tournament I spent weeks trying to cheese it. I threw kitchen sink at it. Nothing stuck. Its unit handling is so good that it deflects attacks even with strategies that just die on paper. It does have a few clear weaknesses, including ones which match your predictions. It has not learned to use Shuttles or Dropships. Human player G5 beat it with Reaver harassment. But Pluto is strong enough that it can often overcome this anyway; Paralyze failed to beat it even while doing significant damage. It also has weak building placement, to the point that it is not competent at Forge/Gateway walls against Zerg, leading it to rely on one-base strategies (and be undertrained against two-base strategies). On the Zerg side it is still somehow strong at busting the walls that it is not trained against. If you are thinking about doing any videos or other content about the topic I'm happy to collaborate.
I did know that overmind is used as a kind of baseline in starcraft AI scene yeah! and wow that's cool. It's quite something to know my sunken colony code survived still today aha :D
I think even the cheater's stolen version can be beat reliably by pros in due time, they just need some time to figure out.
That being said it'll be so fun to have a chat and watch the real pluto in action. And yeah we should do a video. I'll send you a dm 
|
On September 20 2026 05:00 Xeofreestyler wrote: I do wonder if pure self play is the superior choice given the blind spots it has. I remember before AlphaZero was trained with pure self play they first trained AlphaGo on human games and then let it do self-play on top of that to achieve superhuman performance. Given how refined human strats in scbw are I wonder if it could be useful to do the same here.
I'm just wondering if it will ever learn shuttle+reaver or sair usage through self-play alone. I'm assuming pluto learns through small temperature changes so it might take along time for it to stumble upon such strats. So having some human examples might speed up its evolution.
Although admittedly it's hilarious to see it crush mech with m&m in TvT lol
I think alphaGo was initialized with human then later on with RL on top to improve. AlphaZero iirc was entirely from self play.
One issue I see with using human replay for stacraft is that, unlike go, human and AI basically play different games altogether where the AI will have perfect micro while humans do not. Consequently, a lot of human strategies (like muta harass or drops) are attention-attack strategies where you try to spend 1s of time to execute something which would cost the opponent 10s of seconds to resolve. These kind of attention-attack strategies won't work on AI as the pros have found out.
Yet you do need some build orders! I'm a bit skeptical if it'll randomly stumble upon shuttle/reaver or any given strategy, even with a higher temperature in sampling actions. I think higher strategy will just degrade performance rather than promoting diversity. The author said himself that strategies (build orders) had to be loosely guided from human created BOs.
I do think the next form of this bot is a traditional NN controller (as we have now in pluto) combined with a LLM commander that can issue high level commands. This LLM commander can just be any run of the mill LLM like claude or GPT, and humans can basically add some prompt to it for it to "figure out more of the game" so to speak. Ofc engineering such a system requires a lot of work, but I do expect a system like that to emerge in the next 2 years at most.
|
Russian Federation388 Posts
We (human and AI) are already play different games. And the reason is that AI trained to play via different interface. I think what you meant by attention-attack is what we call multi-tasking. Muta harass is not an attention-attack. It's better to call it as abuse of mechanic. It lives in the same space as: fast load/unload reaver, valks patrol micro, stacked lurkers. The reason it doesn't work with AI is much simplier: AI have ability to target low HP muta in the stack. We humans try to do that too, but we only can send command like this: with this pack of rines, target one of the muta in the stack by randomly clicking attack on the stack. And no one would bother to stack lurkers If we could target exact lurker. AlphaZero plays Go via the same interface as human, pick stone -> put stone. Pluto and every other AI in RTS doesn't. We won't see a real move 37 in starcraft, because of the current setup: ability for AI to see a state which human doesn't and to control things human can't reach, indeed it is a different game yet they play.
|
On September 20 2026 18:21 iFU.spx wrote: We (human and AI) are already play different games. And the reason is that AI trained to play via different interface. I think what you meant by attention-attack is what we call multi-tasking. Muta harass is not an attention-attack. It's better to call it as abuse of mechanic. It lives in the same space as: fast load/unload reaver, valks patrol micro, stacked lurkers. The reason it doesn't work with AI is much simplier: AI have ability to target low HP muta in the stack. We humans try to do that too, but we only can send command like this: with this pack of rines, target one of the muta in the stack by randomly clicking attack on the stack. And no one would bother to stack lurkers If we could target exact lurker. AlphaZero plays Go via the same interface as human, pick stone -> put stone. Pluto and every other AI in RTS doesn't. We won't see a real move 37 in starcraft, because of the current setup: ability for AI to see a state which human doesn't and to control things human can't reach, indeed it is a different game yet they play.
Agreed. I think z-ordering (units stacked below another cannot be clicked) is a fundamental game mechanic and ignoring that is very close to cheating. AI should be free from human limits, not game limits. A good litmus test to see what should be allowed or not is: "Can a human with infinite speed and time perform this action?" If the answer is no, then it should be illegal for bots.
|
Belgium6787 Posts
That's a good point, and I think the reason nobody cared until now is that bots were simply nowhere near the strategic thinking level of humans. Perhaps now is a good time for a bwapi v2 where this litmus test is taken into consideration. It does make things considerably harder for developing them.
Something I'm wondering: does a bot know a cloaked unit is nearby? I.e. like how we see an optical distortion
Edit: nvm apparently it's instantly aware of it
|
Watching progamers try to beat a starcraft AI is such a trip. I reread 15+ year old threads theorizing on what AI should do to beat humans. 10 years ago I thought it would take a lot of resources to train such a bot to beat starcraft progamer level just like in Dota2. Now someone can just train a bot easily to beat top players.
It is great that TL has the history.
On October 13 2010 05:23 KwarK wrote:Show nested quote +On October 13 2010 05:13 polgas wrote: Just efficient massing of zerglings/marines/zealots and beat human with perfect micro. Any unit that get injured gets pulled back and healed while the rest of the units keep attacking. This is simply not true. There is a ceiling to how good micro gets and progamers aren't all that far below it. Give Stork 9 dragoons and a computer 8 and Stork will always win, even if the computer has perfect micro. Focus fire + move in cooldown is pretty basic stuff and pulling back wounded units doesn't help you when they die in two volleys and the attacking units have the same movement speed.
|
Belgium6787 Posts
Now someone can just train a bot easily to beat top players.
yeah idk about that, this guy works at meta and probably had access to a shitton of compute
|
On September 21 2026 03:20 Xeofreestyler wrote:yeah idk about that, this guy works at meta and probably had access to a shitton of compute
I doubt the bot can beat top players without high level strategy. there is one thing to "win a game from top players before anyone know you" there's a different thing to "win consistently against top players with a target on your back"
|
Yeah there are just so many obstacles that can trip the bot - new maps, cheese strategies, that I am still impressed that it can still reach this level of being able to beat them.
|
are there any youtube videos of Pluto going up against Overmind style 40,000 apm muta micro zerg bots?
|
On September 21 2026 12:44 Crimson)S(hadow wrote: are there any youtube videos of Pluto going up against Overmind style 40,000 apm muta micro zerg bots? overmind is very weak by today's standard, I think it's mostly used as a test-bed to make sure your bot can run lol (but someone with more knowledge should pitch in
and the logic behind overmind's muta micro is largely hard-coded, except with a few tunable parameters like when to pull back, and when to attack. those parameters are tuned in an RL environment where the muta fought different kidn of units.
so basically pluto is already doing that but better. Pluto is without any hard-coded logic and all the movements learned through RL training onto a neural network
|
On September 21 2026 12:44 Crimson)S(hadow wrote: are there any youtube videos of Pluto going up against Overmind style 40,000 apm muta micro zerg bots? Sadly ~nobody was able to run Overmind for about a decade because it required an ancient version of BWAPI and only worked on the exact map pool from the AIIDE 2010 tournament. It was only this year that I ported it to BWAPI 4.4 and gave it the ability to play on arbitrary maps. It was nice to see it play again in bot Valhalla AKA BASIL.
Pluto just won the 2026 COG StarCraft AI Tournament against traditional rules-based bots (including mine).
McRave has the best current Mutalisk micro among the competitors so check out Pluto's games against him. Easiest way to find them is via the replay browser and easiest way to view them is in your browser with the OpenBW replay viewer.
|
Belgium6787 Posts
On September 22 2026 00:15 PurpleWave wrote:Show nested quote +On September 21 2026 12:44 Crimson)S(hadow wrote: are there any youtube videos of Pluto going up against Overmind style 40,000 apm muta micro zerg bots? Sadly ~nobody was able to run Overmind for about a decade because it required an ancient version of BWAPI and only worked on the exact map pool from the AIIDE 2010 tournament. It was only this year that I ported it to BWAPI 4.4 and gave it the ability to play on arbitrary maps. It was nice to see it play again in bot Valhalla AKA BASIL. Pluto just won the 2026 COG StarCraft AI Tournament against traditional rules-based bots (including mine). McRave has the best current Mutalisk micro among the competitors so check out Pluto's games against him. Easiest way to find them is via the replay browser and easiest way to view them is in your browser with the OpenBW replay viewer.
trying to download some reps like https://davechurchill.ca/starcraft/cog/results/replays/MCRAVE/00018-MCRA_PLUT.REP
but I'm getting 404, any idea?
|
the links to the replays on that page 404 for me
|
|
|
one thing I noticed after watching more of Pluto's playing is that as soon as it attacks, the opponent loses in 30 seconds.
I think the reasons are as follows: 1. pluto is trained in self-play, so it imagined its human opponent to be itself, i.e. with perfect micro 2. pluto maintains some kind of "army value advantage" (let's call it AVA) estimation, to decide if it can beat or not beat the opponent army 3. the online pluto uses maphack, so it has a perfect information to calculate AVA 4. so when pluto commits to an attack, it is imagining that it can beat the perfectly microed opponent army 5. but human's micro is far worse than pluto, so human will lose a chunk of relative army value 6. pluto's AVA goes wayyy up, and goes into a frenzy and human dies
|
On September 22 2026 00:15 PurpleWave wrote:Show nested quote +On September 21 2026 12:44 Crimson)S(hadow wrote: are there any youtube videos of Pluto going up against Overmind style 40,000 apm muta micro zerg bots? Sadly ~nobody was able to run Overmind for about a decade because it required an ancient version of BWAPI and only worked on the exact map pool from the AIIDE 2010 tournament. It was only this year that I ported it to BWAPI 4.4 and gave it the ability to play on arbitrary maps. It was nice to see it play again in bot Valhalla AKA BASIL. Pluto just won the 2026 COG StarCraft AI Tournament against traditional rules-based bots (including mine). McRave has the best current Mutalisk micro among the competitors so check out Pluto's games against him. Easiest way to find them is via the replay browser and easiest way to view them is in your browser with the OpenBW replay viewer.
nice, i'll check out McRave's games, there are plenty in the 3k Pluto reps linked in the other thread, and so far every replay has worked just watching through remastered.
|
It's very nice reading from evan and stryker.
And yes there should be some reasonable limits. Stuff like maphack and UI circumvention techniques is like giving Chess AI and extra move or two.
|
I wonder if the creator is moving away from pure W/L-signal RL + strategy bandit to try to learn efficiently from these human games (which would be a uselessly small sample size otherwise), or from relatively small amounts of high-quality counterfactual-pair training data on things like drops and dark swarm. Given that it's already got a probably pretty well-calibrated win probability/value head, I think fine-tuning it to learn these things could probably be done very cheaply.
|
Russian Federation3632 Posts
The bit about Pluto calculating likes its playing against itself makes me wonder if Pluto having contempt would make it even more effective against human players.
|
|
|
|
|
|