On August 06 2016 11:43 imp42 wrote: The winner of the 2015 SSCAI tournament actually did quite decent macro. That is, expanding quickly and massing hydra (against a two-base protoss turtle attempting to max on carriers, easy 3-0 victory).
> What are the best strategies for a StarCraft bot? > I don't think that the programmers who write bots for the StarCraft AI competitions ever asked this question.
You are mixing two different things here. Competitions like the SSCAI were created as a playing field for Artificial Intelligence students. A good Starcraft bot has probably very little Artificial Intelligence in it, rather it just follows a set of hard-coded rules.
In other words: programmers of AI competitions don't ask themselves that question because a real AI should find out a good strategy and not just execute one it has been fed. That is the whole point of AI. Look at chess or go: programmers define the win condition, implement an optimized algorithm to find it, and let the program do the rest. They don't implement something like "go for a gambit because it is a good strat"
Now to actually answer your question on a more abstract level: What would be a good strat for a Starcraft bot? blah blah HTH
starcraft is a completely different genre compared to chess and go, for all their crazy mechanics the AI is stuck at D level even till now ever wonder why? i bet the programmers who do not even have fundamental competitive understanding of the game will not.
all the crazy marine splitting will not help you if the AI cannot even do something "simple" and intuitive to humans like walling and using mineral walking to defend rushes.
also AI do need to at least start emulating human metagame by adopting standard macro BOs. for example in TvT AI should never ever go bio in midgame (which i shockingly see in the last AI TvT i witness). this is not sc2, marines are weaker, naturally clump together when engaging and no amount of splitting will offset the fact that mech is simply superior here.
the problem with your post is that you actually didn't read/understand the part that you bluntly replaced with "blabla" when quoting me. (hint: tips for creating a good Starcraft bot != tips for creating a good Artificial Intelligence). You don't know who you're talking to, so don't make any wrong assumptions.
...and where are the bots now when it comes to actual results? thats right, Iccup D rank. and you can come up with a lot of theoretical stuff but at the end of the day, where are the results? and this has been going for 6+ years! lmao. if im an investor i would be frightened as hell by this sort of ROI.
just like when programmers are using brute force principles that they use for chess to program weiqi AI and find that they cannot even beat amateur human players 10+ years ago. until google deepmind came up with a novel solution.
its obvious that current methods are not working and perhaps a novel approach is now needed.
it just shows that current theory is most likely inadequate and/or a lot more research needs to be done on how to make a good bot. CURRENT BOTS CANNOT EVEN MAKE A WALL OR DEFEND PROPERLY AGAINST RUSHES, thats like the most fundamental shit. how can you have a good bot when it cannot even grasp the concept and do it in any map? nonsense like 4 pool bots should not even be a threat or figure in the AI metagame at all, yet they are still often falling prey to them. After seeing Berkeley Overmind back in what...2010? so i guess AI development has actually taken a step back? No funding? In essence I don't see any big developments coming anytime soon and it was a real shame Google choose to pick sc2 instead of BW for their Deepmind project.
if you still think im talking nonsense, well i guess you can enjoy having bots continue to languish at D level for the next 5 years or so. they are already there for 6+ years, whats another 5 years? maybe some of the research google did for sc2 can be transferred over to BW, thats a perk i guess.
On August 07 2016 23:03 Simberto wrote:
I am pretty sure that that is not how SC AI works currently. And no, AI does not need to adopt the standard human metagame, if it is unable to play in the way a human can. BW AIs are currently very bad at understanding terrain and making (for humans) very easy decisions. "
then why is letabot constantly asking about how humans respond in certain ingame situatioons OF THE HUMAN METAGAME in the quick questions thread? borrowing BOs of HUMAN METAGAME. if they are really "unable to play in the way a human can." so why bother asking how humans play? hahaahaha. seriously joke post you.
On August 06 2016 11:43 imp42 wrote: The winner of the 2015 SSCAI tournament actually did quite decent macro. That is, expanding quickly and massing hydra (against a two-base protoss turtle attempting to max on carriers, easy 3-0 victory).
> What are the best strategies for a StarCraft bot? > I don't think that the programmers who write bots for the StarCraft AI competitions ever asked this question.
You are mixing two different things here. Competitions like the SSCAI were created as a playing field for Artificial Intelligence students. A good Starcraft bot has probably very little Artificial Intelligence in it, rather it just follows a set of hard-coded rules.
In other words: programmers of AI competitions don't ask themselves that question because a real AI should find out a good strategy and not just execute one it has been fed. That is the whole point of AI. Look at chess or go: programmers define the win condition, implement an optimized algorithm to find it, and let the program do the rest. They don't implement something like "go for a gambit because it is a good strat"
Now to actually answer your question on a more abstract level: What would be a good strat for a Starcraft bot? blah blah HTH
starcraft is a completely different genre compared to chess and go, for all their crazy mechanics the AI is stuck at D level even till now ever wonder why? i bet the programmers who do not even have fundamental competitive understanding of the game will not.
all the crazy marine splitting will not help you if the AI cannot even do something "simple" and intuitive to humans like walling and using mineral walking to defend rushes.
also AI do need to at least start emulating human metagame by adopting standard macro BOs. for example in TvT AI should never ever go bio in midgame (which i shockingly see in the last AI TvT i witness). this is not sc2, marines are weaker, naturally clump together when engaging and no amount of splitting will offset the fact that mech is simply superior here.
the problem with your post is that you actually didn't read/understand the part that you bluntly replaced with "blabla" when quoting me. (hint: tips for creating a good Starcraft bot != tips for creating a good Artificial Intelligence). You don't know who you're talking to, so don't make any wrong assumptions.
...and where are the bots now when it comes to actual results? thats right, Iccup D rank. and you can come up with a lot of theoretical stuff but at the end of the day, where are the results? and this has been going for 6+ years! lmao. if im an investor i would be frightened as hell by this sort of ROI.
just like when programmers are using brute force principles that they use for chess to program weiqi AI and find that they cannot even beat amateur human players 10+ years ago. until google deepmind came up with a novel solution.
its obvious that current methods are not working and perhaps a novel approach is now needed.
it just shows that current theory is most likely inadequate and/or a lot more research needs to be done on how to make a good bot. CURRENT BOTS CANNOT EVEN MAKE A WALL OR DEFEND PROPERLY AGAINST RUSHES, thats like the most fundamental shit. how can you have a good bot when it cannot even grasp the concept and do it in any map? nonsense like 4 pool bots should not even be a threat or figure in the AI metagame at all, yet they are still often falling prey to them. After seeing Berkeley Overmind back in what...2010? so i guess AI development has actually taken a step back? No funding? In essence I don't see any big developments coming anytime soon and it was a real shame Google choose to pick sc2 instead of BW for their Deepmind project.
if you still think im talking nonsense, well i guess you can enjoy having bots continue to languish at D level for the next 5 years or so. they are already there for 6+ years, whats another 5 years? maybe some of the research google did for sc2 can be transferred over to BW, thats a perk i guess.
I am pretty sure that that is not how SC AI works currently. And no, AI does not need to adopt the standard human metagame, if it is unable to play in the way a human can. BW AIs are currently very bad at understanding terrain and making (for humans) very easy decisions. "
then why is letabot constantly asking about how humans respond in certain ingame situatioons OF THE HUMAN METAGAME in the quick questions thread? borrowing BOs of HUMAN METAGAME. if they are really "unable to play in the way a human can." so why bother asking how humans play? hahaahaha. seriously joke post you.
I was under the impression that we were having a polite discussion. Apparently you see this differently and think it is reasonable to be completely impolite instead of formulating an argument. I do not think i want to continue debating with you.
On August 06 2016 11:43 imp42 wrote: The winner of the 2015 SSCAI tournament actually did quite decent macro. That is, expanding quickly and massing hydra (against a two-base protoss turtle attempting to max on carriers, easy 3-0 victory).
> What are the best strategies for a StarCraft bot? > I don't think that the programmers who write bots for the StarCraft AI competitions ever asked this question.
You are mixing two different things here. Competitions like the SSCAI were created as a playing field for Artificial Intelligence students. A good Starcraft bot has probably very little Artificial Intelligence in it, rather it just follows a set of hard-coded rules.
In other words: programmers of AI competitions don't ask themselves that question because a real AI should find out a good strategy and not just execute one it has been fed. That is the whole point of AI. Look at chess or go: programmers define the win condition, implement an optimized algorithm to find it, and let the program do the rest. They don't implement something like "go for a gambit because it is a good strat"
Now to actually answer your question on a more abstract level: What would be a good strat for a Starcraft bot? blah blah HTH
starcraft is a completely different genre compared to chess and go, for all their crazy mechanics the AI is stuck at D level even till now ever wonder why? i bet the programmers who do not even have fundamental competitive understanding of the game will not.
all the crazy marine splitting will not help you if the AI cannot even do something "simple" and intuitive to humans like walling and using mineral walking to defend rushes.
also AI do need to at least start emulating human metagame by adopting standard macro BOs. for example in TvT AI should never ever go bio in midgame (which i shockingly see in the last AI TvT i witness). this is not sc2, marines are weaker, naturally clump together when engaging and no amount of splitting will offset the fact that mech is simply superior here.
the problem with your post is that you actually didn't read/understand the part that you bluntly replaced with "blabla" when quoting me. (hint: tips for creating a good Starcraft bot != tips for creating a good Artificial Intelligence). You don't know who you're talking to, so don't make any wrong assumptions.
...and where are the bots now when it comes to actual results? thats right, Iccup D rank. and you can come up with a lot of theoretical stuff but at the end of the day, where are the results? and this has been going for 6+ years! lmao. if im an investor i would be frightened as hell by this sort of ROI.
just like when programmers are using brute force principles that they use for chess to program weiqi AI and find that they cannot even beat amateur human players 10+ years ago. until google deepmind came up with a novel solution.
its obvious that current methods are not working and perhaps a novel approach is now needed.
it just shows that current theory is most likely inadequate and/or a lot more research needs to be done on how to make a good bot. CURRENT BOTS CANNOT EVEN MAKE A WALL OR DEFEND PROPERLY AGAINST RUSHES, thats like the most fundamental shit. how can you have a good bot when it cannot even grasp the concept and do it in any map? nonsense like 4 pool bots should not even be a threat or figure in the AI metagame at all, yet they are still often falling prey to them. After seeing Berkeley Overmind back in what...2010? so i guess AI development has actually taken a step back? No funding? In essence I don't see any big developments coming anytime soon and it was a real shame Google choose to pick sc2 instead of BW for their Deepmind project.
if you still think im talking nonsense, well i guess you can enjoy having bots continue to languish at D level for the next 5 years or so. they are already there for 6+ years, whats another 5 years? maybe some of the research google did for sc2 can be transferred over to BW, thats a perk i guess.
I am pretty sure that that is not how SC AI works currently. And no, AI does not need to adopt the standard human metagame, if it is unable to play in the way a human can. BW AIs are currently very bad at understanding terrain and making (for humans) very easy decisions. "
then why is letabot constantly asking about how humans respond in certain ingame situatioons OF THE HUMAN METAGAME in the quick questions thread? borrowing BOs of HUMAN METAGAME. if they are really "unable to play in the way a human can." so why bother asking how humans play? hahaahaha. seriously joke post you.
You obviously never seen LetaBot play, because if you did you would know that it was able to build a wall even in early 2014. Anyway watch this video of LetaBot vs fischei ( C+ player):
If that is not a wall at the 1 minute mark then what is it?
Any bot that wants to truly beat a Starcraft player should have serious limitations on APM, micro and control groups to make sure that it havent gained advatage by micro usage but by pure strategy/tactics.
On August 08 2016 22:08 Zedd wrote: Any bot that wants to truly beat a Starcraft player should have serious limitations on APM, micro and control groups to make sure that it havent gained advatage by micro usage but by pure strategy/tactics.
Zedd, I would have agreed if you had said "Artificial Intelligence" instead of "bot".
At the current stage, if you want a bot to win against a human player, you need to abuse mechanisms like APM. Note that a strong bot, especially if working the way I advocate (executing one or more solid timing pushes), does not contain a lot of AI.
In this case, I absolutely agree that a limitation on APM to what is humanly possible makes absolutely sense. That is, limit actions per minute at 300-350 and introduce some delay for "clicks" to simulate the distance on screen the mouse has to cover. For AI vs AI games these limitations do not really matter (although they will alter the metagame). It just makes it much more easy to compare AI decision making to human decision making.
On August 08 2016 10:53 Probemicro wrote: if you still think im talking nonsense, well i guess you can enjoy having bots continue to languish at D level for the next 5 years or so. [...]
The problem is not your lack of knowledge. I think you know more than enough to contribute meaningfully to the discussion. the problem is that you do not seem to appreciate that there are others here, who also might be very qualified. This coupled with a slightly aggressive/ignorant tone is just not as constructive as it could be.
On August 08 2016 22:08 Zedd wrote: Any bot that wants to truly beat a Starcraft player should have serious limitations on APM, micro and control groups to make sure that it havent gained advatage by micro usage but by pure strategy/tactics.
Zedd, I would have agreed if you had said "Artificial Intelligence" instead of "bot".
At the current stage, if you want a bot to win against a human player, you need to abuse mechanisms like APM. Note that a strong bot, especially if working the way I advocate (executing one or more solid timing pushes), does not contain a lot of AI.
In this case, I absolutely agree that a limitation on APM to what is humanly possible makes absolutely sense. That is, limit actions per minute at 300-350 and introduce some delay for "clicks" to simulate the distance on screen the mouse has to cover. For AI vs AI games these limitations do not really matter (although they will alter the metagame). It just makes it much more easy to compare AI decision making to human decision making.
But what is a purpose of a bot that get its advatage by micro?
I think most people creating bots are hoping that eventually, their bot will be semi-intelligent in terms of understanding the game or that it will behave like human so it will be nearly impossible to differentiate between human and such bot for unbiased spectator.
Creating bot that will win by abusing its mechanical skills is like cheating in school. It is small short term gain but you screw yourself even more in long term.
Btw there is many things that one can imagine as a limitation to bot. For example if you create a delay as you mentioned, it would be very hard for bot to do perfect kite or if you limit minimum selection size, then it will be impossible for a bot to do things like avoiding splash damage.
On August 08 2016 10:53 Probemicro wrote: if you still think im talking nonsense, well i guess you can enjoy having bots continue to languish at D level for the next 5 years or so. [...]
The problem is not your lack of knowledge. I think you know more than enough to contribute meaningfully to the discussion. the problem is that you do not seem to appreciate that there are others here, who also might be very qualified. This coupled with a slightly aggressive/ignorant tone is just not as constructive as it could be.
And the main reason that bots continue to be at D level is that bots are currently mostly something that IT students do when writing their masters thesis. That means that there is not a lot of money in it, and the people write a bot over two years, after which they disappear into oblivion.
Quite obviously, this is not the best way to actually do science and push the boundaries.
But what is a purpose of a bot that get its advatage by micro?
I was trying to disentangle "strategy for a bot" from "Artificial Intelligence" in order to better answer the original question, which was "what is a good strategy for a bot?".
The answer to the OP heavily depends on what his intentions are. If he wants to push the boundaries of AI, yes absolutely, don't take advantage of the micro possibilities. But if you just want a challenging bot to practice or win a bot tournament, I would advice differently. That is, the application and consequently imposed limitations affect the recommended strategy.
EPT
