Welcome

Please register for Total War Access to use the forums. If you're an existing user, your forum details will be merged with Total War Access if you register with the same email or username. For more information please read our FAQ’s here.

Categories

Petition to DeepMind, please make AlphaHammer!

2»

Comments

  • RomeoReject#1666RomeoReject#1666 Registered Users Posts: 2,294
    The only - and I mean only - reason I would want to see the resources spent on an adaptive neural network, would be so that mods like GCCM won't just immediately confuse the AI. Beyond that though, seems like using a sledgehammer to tap in a small nail. Will it work? Sure. But it's a huge waste of effort when a simpler solution exists: Just not having confused AI.
  • Seswatha#7633Seswatha#7633 Registered Users Posts: 4,830
    edited January 2021

    The only - and I mean only - reason I would want to see the resources spent on an adaptive neural network, would be so that mods like GCCM won't just immediately confuse the AI. Beyond that though, seems like using a sledgehammer to tap in a small nail. Will it work? Sure. But it's a huge waste of effort when a simpler solution exists: Just not having confused AI.

    AlphaStar's approach is not a solution for this problem actually, as the model would have to be retrained offline to work with a mod, using hardware far too expensive to be even remotely viable (atm).

    There's a number of other approaches they could use in conjunction to make it work but on its own it's not the solution atm.

    Jman5 said:

    It would be very cool to see something as sophisticated as Deepmind take a crack at Warhammer 2 battles. One of the really neat things about this approach that you alluded to is that it's not as encumbered by our own conventional wisdom and beliefs about how to play the game best.

    Also, the nice thing about Total War battles is that it's not as micro-intensive as Starcraft so it becomes more about positioning and decision making.

    It would be utterly fascinating to see how a machine learning system tackles tactics. Would it use vanguard deploy? Would it corner-camp? What spells would it prefer?

    This is exactly the kind of questions I would love to have answered. Does the AI use ethereals to good effect? Does it shun flyers because they are easy to counter in the AI Meta?

    Does it ever go for multiple casters? etc.
    Since TWW is not balanced as well I imagine AI meta being cheese galore. It's also quite possible it would be able to do silly things like dodge 90% of shots from ranged units because it has perfect reflexes and can predict trajectories, unless this is somehow limited to be fair to a human player.

  • WojmirVonCarsteinWojmirVonCarstein Registered Users Posts: 1,598

    The only - and I mean only - reason I would want to see the resources spent on an adaptive neural network, would be so that mods like GCCM won't just immediately confuse the AI. Beyond that though, seems like using a sledgehammer to tap in a small nail. Will it work? Sure. But it's a huge waste of effort when a simpler solution exists: Just not having confused AI.

    AlphaStar's approach is not a solution for this problem actually, as the model would have to be retrained offline to work with a mod, using hardware far too expensive to be even remotely viable (atm).

    There's a number of other approaches they could use in conjunction to make it work but on its own it's not the solution atm.

    Jman5 said:

    It would be very cool to see something as sophisticated as Deepmind take a crack at Warhammer 2 battles. One of the really neat things about this approach that you alluded to is that it's not as encumbered by our own conventional wisdom and beliefs about how to play the game best.

    Also, the nice thing about Total War battles is that it's not as micro-intensive as Starcraft so it becomes more about positioning and decision making.

    It would be utterly fascinating to see how a machine learning system tackles tactics. Would it use vanguard deploy? Would it corner-camp? What spells would it prefer?

    This is exactly the kind of questions I would love to have answered. Does the AI use ethereals to good effect? Does it shun flyers because they are easy to counter in the AI Meta?

    Does it ever go for multiple casters? etc.
    Since TWW is not balanced as well I imagine AI meta being cheese galore. It's also quite possible it would be able to do silly things like dodge 90% of shots from ranged units because it has perfect reflexes and can predict trajectories, unless this is somehow limited to be fair to a human player.
    The issue of AI abusing mechanics (dodging ranged attacks) is similar to the way Alphastar abused Blink stalkers. Microd perfectly, blink stalkers are beyond broken.

    So what was done (and IMO correctly) was that the AI was limited to human like mechanical limitations. So APM were around "pro" level. But even here, The AI was better because it could micro at 500 or 600 APM in crucial moments.

    So presumably, AI would be limited to certain number of actions per second and maybe some delay would be implemented so that the AI could not instantaneously start dodging spells/missiles, but instead there would be some short lag time before it could take such actions.

    But I agree that given that the game is not anywhere close to the balance that Starcraft II has, AI would quickly pick up on cheesy tactics. Honestly this is not dissimilar to human behaviour, though in the hands of the AI the cheese would be exploited to the max. Still it would be cool to see what kind of builds AI decides to use depending on MU and even map
  • WojmirVonCarsteinWojmirVonCarstein Registered Users Posts: 1,598

    The only - and I mean only - reason I would want to see the resources spent on an adaptive neural network, would be so that mods like GCCM won't just immediately confuse the AI. Beyond that though, seems like using a sledgehammer to tap in a small nail. Will it work? Sure. But it's a huge waste of effort when a simpler solution exists: Just not having confused AI.

    Again, I am not suggesting this these resource be actually spent on this. It's not like anyone would listen to me LOL. But it's just a cool thought experiment.

    Also, even if there was some break through in quantum computing and a Desktop could train such AI in minutes to be used in the game, this would still not be something we would want. As a player you want to have the sense of being able to outsmart the AI and win against stacked odds. You definitely don't want to face an AI that can destroy with equal armies because in that case, you would basically need vastly superior armies to even stand a chance. This is not a very fun position to be in.

    For practical AI changes, I started a new thread in the balance section with some ideas.
  • Seswatha#7633Seswatha#7633 Registered Users Posts: 4,830
    edited January 2021

    Jman5 said:

    It would be very cool to see something as sophisticated as Deepmind take a crack at Warhammer 2 battles. One of the really neat things about this approach that you alluded to is that it's not as encumbered by our own conventional wisdom and beliefs about how to play the game best.

    Also, the nice thing about Total War battles is that it's not as micro-intensive as Starcraft so it becomes more about positioning and decision making.

    It would be utterly fascinating to see how a machine learning system tackles tactics. Would it use vanguard deploy? Would it corner-camp? What spells would it prefer?

    This is exactly the kind of questions I would love to have answered. Does the AI use ethereals to good effect? Does it shun flyers because they are easy to counter in the AI Meta?

    Does it ever go for multiple casters? etc.

    The only - and I mean only - reason I would want to see the resources spent on an adaptive neural network, would be so that mods like GCCM won't just immediately confuse the AI. Beyond that though, seems like using a sledgehammer to tap in a small nail. Will it work? Sure. But it's a huge waste of effort when a simpler solution exists: Just not having confused AI.

    AlphaStar's approach is not a solution for this problem actually, as the model would have to be retrained offline to work with a mod, using hardware far too expensive to be even remotely viable (atm).

    There's a number of other approaches they could use in conjunction to make it work but on its own it's not the solution atm.

    Jman5 said:

    It would be very cool to see something as sophisticated as Deepmind take a crack at Warhammer 2 battles. One of the really neat things about this approach that you alluded to is that it's not as encumbered by our own conventional wisdom and beliefs about how to play the game best.

    Also, the nice thing about Total War battles is that it's not as micro-intensive as Starcraft so it becomes more about positioning and decision making.

    It would be utterly fascinating to see how a machine learning system tackles tactics. Would it use vanguard deploy? Would it corner-camp? What spells would it prefer?

    This is exactly the kind of questions I would love to have answered. Does the AI use ethereals to good effect? Does it shun flyers because they are easy to counter in the AI Meta?

    Does it ever go for multiple casters? etc.
    Since TWW is not balanced as well I imagine AI meta being cheese galore. It's also quite possible it would be able to do silly things like dodge 90% of shots from ranged units because it has perfect reflexes and can predict trajectories, unless this is somehow limited to be fair to a human player.
    The issue of AI abusing mechanics (dodging ranged attacks) is similar to the way Alphastar abused Blink stalkers. Microd perfectly, blink stalkers are beyond broken.

    So what was done (and IMO correctly) was that the AI was limited to human like mechanical limitations. So APM were around "pro" level. But even here, The AI was better because it could micro at 500 or 600 APM in crucial moments.

    So presumably, AI would be limited to certain number of actions per second and maybe some delay would be implemented so that the AI could not instantaneously start dodging spells/missiles, but instead there would be some short lag time before it could take such actions.

    But I agree that given that the game is not anywhere close to the balance that Starcraft II has, AI would quickly pick up on cheesy tactics. Honestly this is not dissimilar to human behaviour, though in the hands of the AI the cheese would be exploited to the max. Still it would be cool to see what kind of builds AI decides to use depending on MU and even map
    The cool thing about limiting AlphaStar AI's micro was that it actually had to become a lot smarter when it could no longer win through micro alone.

    In case of TWW though, it would have to be some action lag as you said in addition to the APM limit, tbh I don't think the APM limit would even mean much for TWW as the effective APM to do whatever you want to do when you're managing like 20 units at a time is pretty low in comparison to SC2.

    Even pro level human 600 APM would most likely be already wasted.

  • WojmirVonCarsteinWojmirVonCarstein Registered Users Posts: 1,598

    Jman5 said:

    It would be very cool to see something as sophisticated as Deepmind take a crack at Warhammer 2 battles. One of the really neat things about this approach that you alluded to is that it's not as encumbered by our own conventional wisdom and beliefs about how to play the game best.

    Also, the nice thing about Total War battles is that it's not as micro-intensive as Starcraft so it becomes more about positioning and decision making.

    It would be utterly fascinating to see how a machine learning system tackles tactics. Would it use vanguard deploy? Would it corner-camp? What spells would it prefer?

    This is exactly the kind of questions I would love to have answered. Does the AI use ethereals to good effect? Does it shun flyers because they are easy to counter in the AI Meta?

    Does it ever go for multiple casters? etc.

    The only - and I mean only - reason I would want to see the resources spent on an adaptive neural network, would be so that mods like GCCM won't just immediately confuse the AI. Beyond that though, seems like using a sledgehammer to tap in a small nail. Will it work? Sure. But it's a huge waste of effort when a simpler solution exists: Just not having confused AI.

    AlphaStar's approach is not a solution for this problem actually, as the model would have to be retrained offline to work with a mod, using hardware far too expensive to be even remotely viable (atm).

    There's a number of other approaches they could use in conjunction to make it work but on its own it's not the solution atm.

    Jman5 said:

    It would be very cool to see something as sophisticated as Deepmind take a crack at Warhammer 2 battles. One of the really neat things about this approach that you alluded to is that it's not as encumbered by our own conventional wisdom and beliefs about how to play the game best.

    Also, the nice thing about Total War battles is that it's not as micro-intensive as Starcraft so it becomes more about positioning and decision making.

    It would be utterly fascinating to see how a machine learning system tackles tactics. Would it use vanguard deploy? Would it corner-camp? What spells would it prefer?

    This is exactly the kind of questions I would love to have answered. Does the AI use ethereals to good effect? Does it shun flyers because they are easy to counter in the AI Meta?

    Does it ever go for multiple casters? etc.
    Since TWW is not balanced as well I imagine AI meta being cheese galore. It's also quite possible it would be able to do silly things like dodge 90% of shots from ranged units because it has perfect reflexes and can predict trajectories, unless this is somehow limited to be fair to a human player.
    The issue of AI abusing mechanics (dodging ranged attacks) is similar to the way Alphastar abused Blink stalkers. Microd perfectly, blink stalkers are beyond broken.

    So what was done (and IMO correctly) was that the AI was limited to human like mechanical limitations. So APM were around "pro" level. But even here, The AI was better because it could micro at 500 or 600 APM in crucial moments.

    So presumably, AI would be limited to certain number of actions per second and maybe some delay would be implemented so that the AI could not instantaneously start dodging spells/missiles, but instead there would be some short lag time before it could take such actions.

    But I agree that given that the game is not anywhere close to the balance that Starcraft II has, AI would quickly pick up on cheesy tactics. Honestly this is not dissimilar to human behaviour, though in the hands of the AI the cheese would be exploited to the max. Still it would be cool to see what kind of builds AI decides to use depending on MU and even map
    The cool thing about limiting AlphaStar AI's micro was that it actually had to become a lot smarter when it could no longer win through micro alone.

    In case of TWW though, it would have to be some action lag as you said in addition to the APM limit, tbh I don't think the APM limit would even mean much for TWW as the effective APM to do whatever you want to do when you're managing like 20 units at a time is pretty low in comparison to SC2.

    Even pro level human 600 APM would most likely be already wasted.
    APM itself is not really a good measure as most of those actions are multiple clicks for the same action. Interestingly enough I read that the AI learned to "mimic" this behaviour and it would also "double/triple" click, just like a human would despite there being no reason for the AI to do so.

    What is more important is EAPM (effetive actions per minute) which rarely go above 120.
  • Seswatha#7633Seswatha#7633 Registered Users Posts: 4,830

    Jman5 said:

    It would be very cool to see something as sophisticated as Deepmind take a crack at Warhammer 2 battles. One of the really neat things about this approach that you alluded to is that it's not as encumbered by our own conventional wisdom and beliefs about how to play the game best.

    Also, the nice thing about Total War battles is that it's not as micro-intensive as Starcraft so it becomes more about positioning and decision making.

    It would be utterly fascinating to see how a machine learning system tackles tactics. Would it use vanguard deploy? Would it corner-camp? What spells would it prefer?

    This is exactly the kind of questions I would love to have answered. Does the AI use ethereals to good effect? Does it shun flyers because they are easy to counter in the AI Meta?

    Does it ever go for multiple casters? etc.

    The only - and I mean only - reason I would want to see the resources spent on an adaptive neural network, would be so that mods like GCCM won't just immediately confuse the AI. Beyond that though, seems like using a sledgehammer to tap in a small nail. Will it work? Sure. But it's a huge waste of effort when a simpler solution exists: Just not having confused AI.

    AlphaStar's approach is not a solution for this problem actually, as the model would have to be retrained offline to work with a mod, using hardware far too expensive to be even remotely viable (atm).

    There's a number of other approaches they could use in conjunction to make it work but on its own it's not the solution atm.

    Jman5 said:

    It would be very cool to see something as sophisticated as Deepmind take a crack at Warhammer 2 battles. One of the really neat things about this approach that you alluded to is that it's not as encumbered by our own conventional wisdom and beliefs about how to play the game best.

    Also, the nice thing about Total War battles is that it's not as micro-intensive as Starcraft so it becomes more about positioning and decision making.

    It would be utterly fascinating to see how a machine learning system tackles tactics. Would it use vanguard deploy? Would it corner-camp? What spells would it prefer?

    This is exactly the kind of questions I would love to have answered. Does the AI use ethereals to good effect? Does it shun flyers because they are easy to counter in the AI Meta?

    Does it ever go for multiple casters? etc.
    Since TWW is not balanced as well I imagine AI meta being cheese galore. It's also quite possible it would be able to do silly things like dodge 90% of shots from ranged units because it has perfect reflexes and can predict trajectories, unless this is somehow limited to be fair to a human player.
    The issue of AI abusing mechanics (dodging ranged attacks) is similar to the way Alphastar abused Blink stalkers. Microd perfectly, blink stalkers are beyond broken.

    So what was done (and IMO correctly) was that the AI was limited to human like mechanical limitations. So APM were around "pro" level. But even here, The AI was better because it could micro at 500 or 600 APM in crucial moments.

    So presumably, AI would be limited to certain number of actions per second and maybe some delay would be implemented so that the AI could not instantaneously start dodging spells/missiles, but instead there would be some short lag time before it could take such actions.

    But I agree that given that the game is not anywhere close to the balance that Starcraft II has, AI would quickly pick up on cheesy tactics. Honestly this is not dissimilar to human behaviour, though in the hands of the AI the cheese would be exploited to the max. Still it would be cool to see what kind of builds AI decides to use depending on MU and even map
    The cool thing about limiting AlphaStar AI's micro was that it actually had to become a lot smarter when it could no longer win through micro alone.

    In case of TWW though, it would have to be some action lag as you said in addition to the APM limit, tbh I don't think the APM limit would even mean much for TWW as the effective APM to do whatever you want to do when you're managing like 20 units at a time is pretty low in comparison to SC2.

    Even pro level human 600 APM would most likely be already wasted.
    APM itself is not really a good measure as most of those actions are multiple clicks for the same action. Interestingly enough I read that the AI learned to "mimic" this behaviour and it would also "double/triple" click, just like a human would despite there being no reason for the AI to do so.

    What is more important is EAPM (effetive actions per minute) which rarely go above 120.
    EPM is indeed lower, though for pros it can be much higher than 120 e.g 250-300 (at least using SC2 method of claculation). But the point was that for TWW both APM and EPM required to play it optimally would be a lot lower than for SC2 where even pro level APM is not enough to do things an AI would be capable of.

  • Xenos777Xenos777 Registered Users Posts: 8,038
    It would somewhat cool, but most of the strategy in 4X games (and in actual war, I would say) is about forcing the enemy to fight battles you've already won in principle. So it wouldn't really change much in campaign.
  • TheShiroOfDaltonTheShiroOfDalton Registered Users Posts: 34,001
    I have to bring up the issue of naval battles in R2 again. Naval battles in R2 were designed in a way that the AI's ability to effectively micro 20-40 units at once really came to the forefront and unless you used pause and slow-motion a lot, a battle of full fleets on an open sea map would be the most hectic battle ever since ships were both fast and could die very quickly. Hence why R2's naval battles quickly got a very bad reputation. That they were buggy and ships could be very unresponsive to player orders didn't help.

    When CA_Sofia patched the game, they massively reduced the AI's ability to abuse its faster input speed and that made naval battles a heckuva lot more manageable. They also did away with the AI parking small ships right next to yours and then making micro rams over and over again as a bonus.

    This shows an AI that can outplay the player consistently isn't the berries, it's in fact a reason why it shouldn't be designed that way. If it was given sufficient meta-knowledge then you'd be meeting nothing but buffstacked ranged stacks and with perfect micro they'd be unbelievably annoying to fight. I mean range-heavy WE or Skaven builds are already excruciating to deal with, imagine that combined with an AI that will always be on point with its unit usage.
  • WojmirVonCarsteinWojmirVonCarstein Registered Users Posts: 1,598
    Xenos7777 said:

    It would somewhat cool, but most of the strategy in 4X games (and in actual war, I would say) is about forcing the enemy to fight battles you've already won in principle. So it wouldn't really change much in campaign.

    My idea would not be to have this in campaign as it would make campaign basically unplayable.
  • Seswatha#7633Seswatha#7633 Registered Users Posts: 4,830
    edited January 2021

    I have to bring up the issue of naval battles in R2 again. Naval battles in R2 were designed in a way that the AI's ability to effectively micro 20-40 units at once really came to the forefront and unless you used pause and slow-motion a lot, a battle of full fleets on an open sea map would be the most hectic battle ever since ships were both fast and could die very quickly. Hence why R2's naval battles quickly got a very bad reputation. That they were buggy and ships could be very unresponsive to player orders didn't help.

    When CA_Sofia patched the game, they massively reduced the AI's ability to abuse its faster input speed and that made naval battles a heckuva lot more manageable. They also did away with the AI parking small ships right next to yours and then making micro rams over and over again as a bonus.

    This shows an AI that can outplay the player consistently isn't the berries, it's in fact a reason why it shouldn't be designed that way. If it was given sufficient meta-knowledge then you'd be meeting nothing but buffstacked ranged stacks and with perfect micro they'd be unbelievably annoying to fight. I mean range-heavy WE or Skaven builds are already excruciating to deal with, imagine that combined with an AI that will always be on point with its unit usage.

    You could also argue that a competent AI would mean the devs would need to finally balance the campaign as well and not just MP. Whether it will be worth the effort is another question.

    And yes, superhuman reaction speeds would to be limited to something that's comparable to an average player for a SP AI. This was already done with AlphaStar, though it's still faster than most pros I think.

  • WojmirVonCarsteinWojmirVonCarstein Registered Users Posts: 1,598

    I have to bring up the issue of naval battles in R2 again. Naval battles in R2 were designed in a way that the AI's ability to effectively micro 20-40 units at once really came to the forefront and unless you used pause and slow-motion a lot, a battle of full fleets on an open sea map would be the most hectic battle ever since ships were both fast and could die very quickly. Hence why R2's naval battles quickly got a very bad reputation. That they were buggy and ships could be very unresponsive to player orders didn't help.

    When CA_Sofia patched the game, they massively reduced the AI's ability to abuse its faster input speed and that made naval battles a heckuva lot more manageable. They also did away with the AI parking small ships right next to yours and then making micro rams over and over again as a bonus.

    This shows an AI that can outplay the player consistently isn't the berries, it's in fact a reason why it shouldn't be designed that way. If it was given sufficient meta-knowledge then you'd be meeting nothing but buffstacked ranged stacks and with perfect micro they'd be unbelievably annoying to fight. I mean range-heavy WE or Skaven builds are already excruciating to deal with, imagine that combined with an AI that will always be on point with its unit usage.

    You could also argue that a competent AI would mean the devs would need to finally balance the campaign as well and not just MP. Whether it will be worth the effort is another question.

    And yes, superhuman reaction speeds would to be limited to something that's comparable to an average player for a SP AI. This was already done with AlphaStar, though it's still faster than most pros I think.
    I would say reaction times should be LOWER than average player given that AI will never "not see" something like a player would.

    Alpha star does have "pro level" APM. However, there are still peaks it hits in crucial moments that are basically unatainable for a human. Also, Alphastar will never "mislick" the wrong target or area to move to like a human might.
  • TheShiroOfDaltonTheShiroOfDalton Registered Users Posts: 34,001

    Xenos7777 said:

    It would somewhat cool, but most of the strategy in 4X games (and in actual war, I would say) is about forcing the enemy to fight battles you've already won in principle. So it wouldn't really change much in campaign.

    My idea would not be to have this in campaign as it would make campaign basically unplayable.
    Not to speak that this is exactly how the WH1 AI acted, it would only take battles when it had an overwhelming advantage and it was plain annoying to have the AI otherwise run from any fight or camp in settlements.
Sign In or Register to comment.