Petition to DeepMind, please make AlphaHammer!

WojmirVonCarstein

For those not aware, Google's DeepMind worked on an AI for Starcraft II called "AlphaStar". Essentially, they used the most sophisticated Machine Learning techniques to have an AI "teach" itself to play StarCraft II without cheating. After some initial complaints about the AI cheating by being able to "see" the entire map on it's "screen" and being able to take unhuman like number of actions per second, the developers limited both and Alphastar was limited to human like behaviour and information. It was able to play on Ranked ladder and achieve Grandmaster status with all 3 races. It's games are fascinating to watch for anyone interested in StarCraft II

Wouldn't it be cool if they did this for Warhammer for Quick Battles specifically? I was wondering how much more (less?) of a challenge would it be? Starcraft has 3 different races thus only 3^2 = 9 possible MUs. While Warhammer has 15 unique factions, which will mean 225 unique MUs. Not to mention sub-factions that have access to unique units (Clan Angrund, Avelorn, Chevaliers de Lyonesse, Followers of Nagash, Sartosa, The Drowned).

The AI would first have to learn to build armies and then of course play out the battles. It would be so exciting to see the AI do this. Would we see builds that we have never seen before? Strategies that humans have overlooked? Units that we once thought UP become suddenly OP in the hands of the AI. Which spells/items would the AI choose?

Of course, some limitation would have to put on the AI for it not to be "super human". For example dodging would have to be somehow limited, otherwise ranged units would not be able to hit anything as the AI could calculate trajectories perfectly and evade all damage, especially with fast units. Bombardment, Vortex, Wind spells would also become close to useless. So either limit actions per second to be similar to top human players or make sure there is some RNG involved in how fast it can react to spells/ranged attacks.

The AI would also have to be limited in it's use of exploits such as White line/corner camping, forest camping, cycle charging in end game etc. So maybe some sort of tournament rules would have to be applied so that it always abides by these?

Thoughts?

Crossil#2134

Because Starcraft 2 is an Esports game. Highly competitive. It's also why there was one for, what was it, LoL or Dota?

TWW ain't.

WojmirVonCarstein

Crossil said:

Because Starcraft 2 is an Esports game. Highly competitive. It's also why there was one for, what was it, LoL or Dota?

TWW ain't.

I understand this, I was just thinking of how cool it would be and wanted to discuss to see if people thing that AlphaHammer would be able to be competitive, would it be OP, are there any factions/strategies/builds it would favour over others, how it would approach MUs, how would map alter its strategies etc.

UberReptilian#8431

I'd rather just CA make the A.I semi-intelligent. It's all the A without the I in this game.

Crossil#2134

WojmirVonCarstein said:

Crossil said:

Because Starcraft 2 is an Esports game. Highly competitive. It's also why there was one for, what was it, LoL or Dota?

TWW ain't.

I understand this, I was just thinking of how cool it would be and wanted to discuss to see if people thing that AlphaHammer would be able to be competitive, would it be OP, are there any factions/strategies/builds it would favour over others, how it would approach MUs, how would map alter its strategies etc.

I don't know. Having to deal with so many battles is already kinda annoying but getting absolutely butchered by the AI when I'm just trying to have a relaxing game is not really on my priorities list.

I wouldn't mind what UberReptilian says.

WojmirVonCarstein

Crossil said:

WojmirVonCarstein said:

Crossil said:

Because Starcraft 2 is an Esports game. Highly competitive. It's also why there was one for, what was it, LoL or Dota?

TWW ain't.

I understand this, I was just thinking of how cool it would be and wanted to discuss to see if people thing that AlphaHammer would be able to be competitive, would it be OP, are there any factions/strategies/builds it would favour over others, how it would approach MUs, how would map alter its strategies etc.

I don't know. Having to deal with so many battles is already kinda annoying but getting absolutely butchered by the AI when I'm just trying to have a relaxing game is not really on my priorities list.

I wouldn't mind what UberReptilian says.

Again, my intention would not be for that AI to be in the SP game (this was not the case for Starcraft either). It would just be a "player" on the QB ladder that you might bump into once in several dozen games.

WojmirVonCarstein

UberReptilian said:

I'd rather just CA make the A.I semi-intelligent. It's all the A without the I in this game.

This of course is what a lot of people want, the AlphaHammer is just a fun idea to discuss.

snnnffrknlin#6194

This game definitely needs better AI, but sadly there will not be an Alphastar for Total war(yet). It's too expensive to make and maintain. We are talking about millions of dollars just to rent the TPUs necessary to make the AI.

I reccomend reading about Alphastar on their site or just watch this video(from 6:25).

https://youtu.be/84VVFxHAdSY?t=385

UberReptilian#8431

WojmirVonCarstein said:

UberReptilian said:

I'd rather just CA make the A.I semi-intelligent. It's all the A without the I in this game.

This of course is what a lot of people want, the AlphaHammer is just a fun idea to discuss.

I view it more as a pipe dream, but it is an interesting idea. I'd never heard of it anyway. Some basic A.I compentency would be all I need to be satisfied though.

Conanthelibrarian#3850

UberReptilian said:

I'd rather just CA make the A.I semi-intelligent. It's all the A without the I in this game.

I cant imagine how hard it must be, since no one seems to be able to do it.

But I think a few simple changes could make a huge improvement in TW AI:

1. Magic: Mostly use buffs/debuffs, since the AI wastes most damage spells. Or make them only use damage spells when it will hit two or more formations. But stick to buffing and debuffing the main infantry lines as they meet. Damage spells on ARTY would also work,

2. Combat: Have the AI focus on taking out your ARTY first, then archers when possible. The AI rarely wins a battle, but that would slow down the Human player if they were losing units or taking heavy losses and needed to spend turns replenishing.

3. Combat: Only throw their heros and lord into battle against other lords/heros, or wait until they are needed at the battle turns against them. They waste heros and lords too early charging spearmen. And for Races like VC and TK, the main Lord never initiates combat to prevent early crumbling.

4. ARTY: When the AI has an Arty advantage, or even tie, it never attacks and trades arty fire with you until it runs out of ammo or you attack it. It may not win, but it will do serious damage to your units, again, slowing down the players campaign progress. Arty is the easiest thing for the AI to use to do damage, and too often it attacks and spends too much time moving their arty around instead of attacking and doing damage with it.
With that said, the AI should focus on having 2-4 Arty in all its armies for factions that can have it. Or 1 arty per 4 units as it builds an army. Thata way it can hurt the human player and slow them down.

Seswatha#7633

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

Uagrim#4644

Its a huge investment of resources that would be better used elsewhere.

Plus quick battle data is comparatively small and the usable ones are even smaller (draw kiting, corner camping, etc). So pretty much only tournaments which all are very divided widely between patches. Unlike SC2 which saw its last Multiplayer content drop in 2015 and only minor changes since then, TWW gets a bunch of new units dropped every half a year.

It wouldn't even be useful in campaign since Quick Battle is a One time thing, loses do not have strategic consequences and all that matters is that you win the battle.

TheShiroOfDalton

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

Seswatha#7633

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

TheShiroOfDalton

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

The point is that all decisions in TW have some sort of inbuilt input lag, so there's more thinking ahead required whereas in SC or AoE this factor is strongly diminished thanks to the weightless nature of the units on the map.

Seswatha#7633

ShiroAmakusa75 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

The point is that all decisions in TW have some sort of inbuilt input lag, so there's more thinking ahead required whereas in SC or AoE this factor is strongly diminished thanks to the weightless nature of the units on the map.

Committing to a bad engagement is just as bad in SC2 as in TWW, the main difference is that SC2 is more deterministic so you can theoretically calculate the outcome and in TWW outcome will be probabilistic. Though in practice, given fog of war and incomplete information as well as high impact of micro it's not deterministic in SC2 either.

In any case reinforcement learning which is used by AlphaStar in particular doesn't work this way.

davedave1124#4773

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Seswatha#7633

davedave1124 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Old AIs indeed could use a fairly straightforward scripted behavior for base building which is why they kinda suck (but it can be "good enough").

AlphaStar uses reinforcement learning, you can read about high level problem and approach here: https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

But you don't really need to know how they both work to tell that decision space in SC2 is just massively bigger. As they mention they have 10^26 legal moves at any point. Working with limited information and long term planning are also much less of a thing in TWW.

davedave1124#4773

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Old AIs indeed could use a fairly straightforward scripted behavior for base building which is why they kinda suck (but it can be "good enough").

AlphaStar uses reinforcement learning, you can read about high level problem and approach here: https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

But you don't really need to know how they both work to tell that decision space in SC2 is just massively bigger. As they mention they have 10^26 legal moves at any point. Working with limited information and long term planning are also much less of a thing in TWW.

Oh, I think we almost certainly do need to know how each AI works and how useful are they for each game type. As the video mentions above, AlphaStar has serious limits and I've not been shown any evidence it would work for a more complex game like modern TW, I think there's a reason why they used it on a pretty basic RTS game.

Until it's actually tested on a more modern game like TW (not forgetting no one can replicate CA's programming - the reason they have no competition) I'm not going to accept it's some ultimate solve for TW's issues. If it's enough for you - cool.

Seswatha#7633

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Old AIs indeed could use a fairly straightforward scripted behavior for base building which is why they kinda suck (but it can be "good enough").

AlphaStar uses reinforcement learning, you can read about high level problem and approach here: https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

But you don't really need to know how they both work to tell that decision space in SC2 is just massively bigger. As they mention they have 10^26 legal moves at any point. Working with limited information and long term planning are also much less of a thing in TWW.

Oh, I think we almost certainly do need to know how each AI works and how useful are they for each game type. As the video mentions above, AlphaStar has serious limits and I've not been shown any evidence it would work for a more complex game like modern TW, I think there's a reason why they used it on a pretty basic RTS game.

Until it's actually tested on a more modern game like TW (not forgetting no one can replicate CA's programming - the reason they have no competition) I'm not going to accept it's some ultimate solve for TW's issues. If it's enough for you - cool.

TW is not a more complex game in terms of battle AI though, the only thing more complex about it is the number of factions and units, but fundamentally it's simpler. There's also a ton of publicly available information and research on reinforcement learning and its applications in gaming as well as high level approaches used in AlphaStar if you're interested.

The problem with AlphaStar is not that this approach won't work for TWW, it's that it's too expensive.

But everyone's entitled to their opinion of course.

saweendra#3399

Crossil said:

WojmirVonCarstein said:

Crossil said:

Because Starcraft 2 is an Esports game. Highly competitive. It's also why there was one for, what was it, LoL or Dota?

TWW ain't.

I understand this, I was just thinking of how cool it would be and wanted to discuss to see if people thing that AlphaHammer would be able to be competitive, would it be OP, are there any factions/strategies/builds it would favour over others, how it would approach MUs, how would map alter its strategies etc.

I don't know. Having to deal with so many battles is already kinda annoying but getting absolutely butchered by the AI when I'm just trying to have a relaxing game is not really on my priorities list.

I wouldn't mind what UberReptilian says.

Oh they can adjust how hard the ai play you in difficultly settings.

Also they should balance the campaign so its a challenge not annoying.

Bring me to war i want the best possible AI.

WojmirVonCarstein

saweendra said:

Crossil said:

WojmirVonCarstein said:

Crossil said:

Because Starcraft 2 is an Esports game. Highly competitive. It's also why there was one for, what was it, LoL or Dota?

TWW ain't.

I understand this, I was just thinking of how cool it would be and wanted to discuss to see if people thing that AlphaHammer would be able to be competitive, would it be OP, are there any factions/strategies/builds it would favour over others, how it would approach MUs, how would map alter its strategies etc.

I don't know. Having to deal with so many battles is already kinda annoying but getting absolutely butchered by the AI when I'm just trying to have a relaxing game is not really on my priorities list.

I wouldn't mind what UberReptilian says.

Oh they can adjust how hard the ai play you in difficultly settings.

Also they should balance the campaign so its a challenge not annoying.

Bring me to war i want the best possible AI.

This sounds fun, but if you think about it, it would not be. The best possible AI would be able to wreck 99.9% of players in an even odds battle. It would not be fun to have to have vastly superior forces to even have a chance at winning. It's more fun (IMO) to have the inferior force and yet come out on top due to superior tactics during battle, not the other way around

A better AI though is a must.

davedave1124#4773

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Old AIs indeed could use a fairly straightforward scripted behavior for base building which is why they kinda suck (but it can be "good enough").

AlphaStar uses reinforcement learning, you can read about high level problem and approach here: https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

But you don't really need to know how they both work to tell that decision space in SC2 is just massively bigger. As they mention they have 10^26 legal moves at any point. Working with limited information and long term planning are also much less of a thing in TWW.

Oh, I think we almost certainly do need to know how each AI works and how useful are they for each game type. As the video mentions above, AlphaStar has serious limits and I've not been shown any evidence it would work for a more complex game like modern TW, I think there's a reason why they used it on a pretty basic RTS game.

Until it's actually tested on a more modern game like TW (not forgetting no one can replicate CA's programming - the reason they have no competition) I'm not going to accept it's some ultimate solve for TW's issues. If it's enough for you - cool.

TW is not a more complex game in terms of battle AI though, the only thing complex about it is the number of factions and units, but fundamentally it's a simpler.

But everyone's entitled to their opinion of course.

Is it? Any references? Here's the thoughts of a person with a doctorate in AI, who understands TW AI is cutting edge despite what the 'forum experts' say. Opinions are great, as long as they are backed up with something.

https://www.youtube.com/watch?v=XBzTJOYgW0M

Simple games like the above or AoE/K use simple AI because that's all is required for them.

As mentioned above, AlphaStar has clear limits as it only seems able to learn the most efficient way to win on one particular game and I've not seen it perform on any other. AlphaStar also isn't being used to make games 'better' or more enjoyable, it's there to test the quality of AI to do certain tasks, as in efficiency.

Would it make TW a better game? Dunno - more **evidence** required.

davedave1124#4773

saweendra said:

Crossil said:

WojmirVonCarstein said:

Crossil said:

Because Starcraft 2 is an Esports game. Highly competitive. It's also why there was one for, what was it, LoL or Dota?

TWW ain't.

I understand this, I was just thinking of how cool it would be and wanted to discuss to see if people thing that AlphaHammer would be able to be competitive, would it be OP, are there any factions/strategies/builds it would favour over others, how it would approach MUs, how would map alter its strategies etc.

I don't know. Having to deal with so many battles is already kinda annoying but getting absolutely butchered by the AI when I'm just trying to have a relaxing game is not really on my priorities list.

I wouldn't mind what UberReptilian says.

Oh they can adjust how hard the ai play you in difficultly settings.

Also they should balance the campaign so its a challenge not annoying.

Bring me to war i want the best possible AI.

That's the issue, it's very difficult to 'tweek' machine learning programmes, this is a little harder than some people are making out.

Seswatha#7633

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Old AIs indeed could use a fairly straightforward scripted behavior for base building which is why they kinda suck (but it can be "good enough").

AlphaStar uses reinforcement learning, you can read about high level problem and approach here: https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

But you don't really need to know how they both work to tell that decision space in SC2 is just massively bigger. As they mention they have 10^26 legal moves at any point. Working with limited information and long term planning are also much less of a thing in TWW.

Oh, I think we almost certainly do need to know how each AI works and how useful are they for each game type. As the video mentions above, AlphaStar has serious limits and I've not been shown any evidence it would work for a more complex game like modern TW, I think there's a reason why they used it on a pretty basic RTS game.

Until it's actually tested on a more modern game like TW (not forgetting no one can replicate CA's programming - the reason they have no competition) I'm not going to accept it's some ultimate solve for TW's issues. If it's enough for you - cool.

TW is not a more complex game in terms of battle AI though, the only thing complex about it is the number of factions and units, but fundamentally it's a simpler.

But everyone's entitled to their opinion of course.

Is it? Any references? Here's the thoughts of a person with a doctorate in AI, who understands TW AI is cutting edge despite what the 'forum experts' say. Opinions are great, as long as they are backed up with something.

https://www.youtube.com/watch?v=XBzTJOYgW0M

Simple games like the above or AoE/K use simple AI because that's all is required for them.

As mentioned above, AlphaStar has clear limits as it only seems able to learn the most efficient way to win on one particular game and I've not seen it perform on any other. AlphaStar also isn't being used to make games 'better' or more enjoyable, it's there to test the quality of AI to do certain tasks, as in efficiency.

Would it make TW a better game? Dunno - more **evidence** required.

AlphaStar was a project made by the same team that made AlphaGo, which beats the best players in the world at Go, one of the most complex classical turn based strategy games due to huge decision space. It uses a lot of the same core concepts. If you understand how neural networks and reinforcement learning work, you could see how it can be applied to basically any game or a lot of things outside gaming as well. And it's not hard to see that the number of inputs and decision space in TWW battles is a lot smaller than in SC2. But to understand that, you need to understand how these things work so the only reference I can give you is this https://www.coursera.org/specializations/reinforcement-learning but it's a lot of time investment and you'll have to learn basic ML first. Other than that no, I don't have a video of Andrew NG saying that Starcraft is more complex than TWW because he probably doesn't even know about TWW.

But tbh you don't even have to know anything about ML and just play both SC2 and TWW at a decent level to tell that TWW battles are not nearly as complex as SC2.

As for making a better game - it's a separate question, players might not find playing against an AI that always outplays them enjoyable.

davedave1124#4773

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Old AIs indeed could use a fairly straightforward scripted behavior for base building which is why they kinda suck (but it can be "good enough").

AlphaStar uses reinforcement learning, you can read about high level problem and approach here: https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

But you don't really need to know how they both work to tell that decision space in SC2 is just massively bigger. As they mention they have 10^26 legal moves at any point. Working with limited information and long term planning are also much less of a thing in TWW.

Oh, I think we almost certainly do need to know how each AI works and how useful are they for each game type. As the video mentions above, AlphaStar has serious limits and I've not been shown any evidence it would work for a more complex game like modern TW, I think there's a reason why they used it on a pretty basic RTS game.

Until it's actually tested on a more modern game like TW (not forgetting no one can replicate CA's programming - the reason they have no competition) I'm not going to accept it's some ultimate solve for TW's issues. If it's enough for you - cool.

TW is not a more complex game in terms of battle AI though, the only thing complex about it is the number of factions and units, but fundamentally it's a simpler.

But everyone's entitled to their opinion of course.

Is it? Any references? Here's the thoughts of a person with a doctorate in AI, who understands TW AI is cutting edge despite what the 'forum experts' say. Opinions are great, as long as they are backed up with something.

https://www.youtube.com/watch?v=XBzTJOYgW0M

Simple games like the above or AoE/K use simple AI because that's all is required for them.

As mentioned above, AlphaStar has clear limits as it only seems able to learn the most efficient way to win on one particular game and I've not seen it perform on any other. AlphaStar also isn't being used to make games 'better' or more enjoyable, it's there to test the quality of AI to do certain tasks, as in efficiency.

Would it make TW a better game? Dunno - more **evidence** required.

AlphaStar was a project made by the same team that made AlphaGo, which beats the best players in the world at Go, one of the most complex classical turn based strategy games due to huge decision space. If you understand how neural networks and reinforcement learning work, you could see how it can be applied to basically any game or a lot of things outside gaming as well. And it's not hard to see that the number of inputs and decision space in TWW battles is a lot smaller than in SC2. But to understand that, you need to understand how these things work so the only reference I can give you is this https://www.coursera.org/specializations/reinforcement-learning but it's a lot of time investment and you'll have to learn basic ML first. Other than that no, I don't have a video of Andrew NG saying that Starcraft is more complex than TWW because he probably doesn't even know about TWW.

As for making a better game - it's a separate question, players might not find playing against an AI that always outplays them enjoyable.

Games like Go are less complex than basic RTS games like AoKs and Starcraft 2, so let's not take the conversation backwards. Using the idea that the programme is more complex because it beat something less complex tells me you know less about the subject than you think.

Again, you've not shown any evidence that battles are more complex in the 11 year old game SC2 so I'm not going to reply on that until you do.

Again, you missed the point from the expert opinion above that machine learning can only follow simple rules like win in the most efficient way. It cannot be tweeked in a meaningful way so would be completely useless for game design.

Yes, you have no evidence that a basic RTS is more complex than an up to date TW game - that's my point and why I'm rather confused you seem certain about it. To give a reasonable argument I assume you would need to count up the unit types, the effects of positioning, aura abilities, constant abilities, possible combos, use of magic, charge, leadership, elevation, fatigue etc.

What has what and what doesn't and how complex is it? Questions like that around a coherent argument.

Seswatha#7633

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Old AIs indeed could use a fairly straightforward scripted behavior for base building which is why they kinda suck (but it can be "good enough").

AlphaStar uses reinforcement learning, you can read about high level problem and approach here: https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

But you don't really need to know how they both work to tell that decision space in SC2 is just massively bigger. As they mention they have 10^26 legal moves at any point. Working with limited information and long term planning are also much less of a thing in TWW.

Oh, I think we almost certainly do need to know how each AI works and how useful are they for each game type. As the video mentions above, AlphaStar has serious limits and I've not been shown any evidence it would work for a more complex game like modern TW, I think there's a reason why they used it on a pretty basic RTS game.

Until it's actually tested on a more modern game like TW (not forgetting no one can replicate CA's programming - the reason they have no competition) I'm not going to accept it's some ultimate solve for TW's issues. If it's enough for you - cool.

TW is not a more complex game in terms of battle AI though, the only thing complex about it is the number of factions and units, but fundamentally it's a simpler.

But everyone's entitled to their opinion of course.

Is it? Any references? Here's the thoughts of a person with a doctorate in AI, who understands TW AI is cutting edge despite what the 'forum experts' say. Opinions are great, as long as they are backed up with something.

https://www.youtube.com/watch?v=XBzTJOYgW0M

Simple games like the above or AoE/K use simple AI because that's all is required for them.

As mentioned above, AlphaStar has clear limits as it only seems able to learn the most efficient way to win on one particular game and I've not seen it perform on any other. AlphaStar also isn't being used to make games 'better' or more enjoyable, it's there to test the quality of AI to do certain tasks, as in efficiency.

Would it make TW a better game? Dunno - more **evidence** required.

AlphaStar was a project made by the same team that made AlphaGo, which beats the best players in the world at Go, one of the most complex classical turn based strategy games due to huge decision space. If you understand how neural networks and reinforcement learning work, you could see how it can be applied to basically any game or a lot of things outside gaming as well. And it's not hard to see that the number of inputs and decision space in TWW battles is a lot smaller than in SC2. But to understand that, you need to understand how these things work so the only reference I can give you is this https://www.coursera.org/specializations/reinforcement-learning but it's a lot of time investment and you'll have to learn basic ML first. Other than that no, I don't have a video of Andrew NG saying that Starcraft is more complex than TWW because he probably doesn't even know about TWW.

As for making a better game - it's a separate question, players might not find playing against an AI that always outplays them enjoyable.

Games like Go are less complex than basic RTS games like AoKs and Starcraft 2, so let's not take the conversation backwards. Using the idea that the programme is more complex because it beat something less complex tells me you know less about the subject than you think.

Again, you've not shown any evidence that battles are more complex in the 11 year old game SC2 so I'm not going to reply on that until you do.

Again, you missed the point from the expert opinion above that machine learning can only follow simple rules like win in the most efficient way. It cannot be tweeked in a meaningful way so would be completely useless for game design.

Yes, you have no evidence that a basic RTS is more complex than an up to date TW game - that's my point and why I'm rather confused you seem certain about it. To give a reasonable argument I assume you would need to count up the unit types, the effects of positioning, aura abilities, constant abilities, possible combos, use of magic, charge, leadership, elevation, fatigue etc.

What has what and what doesn't and how complex is it? Questions like that around a coherent argument.

I used Go as an example to show that these approaches are applicable universally, not as a more complex game. Obviously they moved from less complex Go to more complex SC2.

And quoting my edit from the previous post "But tbh you don't even have to know anything about ML and just play both SC2 and TWW at a decent level to tell that TWW battles are not nearly as complex as SC2."

SC2 has (compared to TWW battles):
-strategic layer and long term planning
-scouting and working with incomplete information (to a much greater extent), mind games
-just more **** going on with a lot more units, buildings and terrain features, and the number of possible actions you can take at any moment is just a lot higher
-micro of individual units it a lot more complex and many abilities (e.g. force fields) are more complex to use and much less point and click

AlphaStar models are different models trained on a specific MU (e.g. Protoss vs Protoss), so the fact that TWW has more units and factions mostly means they would need more models, but it's just quantity, it's not fundamentally more difficult. And each model would be simpler due to reasons above.

Anyway, if you're really interested in the topic I suggest you reading up on the subject, I understand that not everyone wants to or has the time to learn ML but there's enough popular articles to at least get a high level understanding.

If you're here just to argue let's just agree to disagree.

Jman5#8318

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?

davedave1124#4773

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

davedave1124 said:

MadDemiurg said:

ShiroAmakusa75 said:

MadDemiurg said:

Maybe in like 10 years, AlphaStar models costed millions to just train, that's without the engineering costs, I doubt they'll invest this much money here as TWW is not nearly as famous. If there's a breakthrough in the future to make it a lot cheaper, maybe.

TWW battles are also much less complex than SC2 games and thus less of a challenge for Deepmind to tackle, though the number of possible MU permutations is very large compared to SC2 so that will affect training time.

TWW battles are actually a lot more complex than SC2 games because they bother to actually simulate physics instead of having all units move like a weightless school of fish on maps that have only three fixed heights (ground floor, up floor and flying). In SC2 if you send Zealots into combat they'll march right next to the target and simply start attacking all at the same speed with the same damage output, in TW you have charge speed and impact plus charge bonus to keep track of. Sure you could argue that SC2 players need to also keep track of their base and economy, but that's more a question of build orders and raw APM.

That's the complexity of the simulation, not the complexity of the decision making. Decision making in SC2 is a lot more complex as it involves base building and management, high level strategy and transitions, scouting and working with incomplete information and many other things. Not to mention each individual unit needs to be actually microed by the AI be it using abilities like blink or force field or just doing basic stutter step micro, harassing, doing drops etc. The amount of decisions and actions AI has to do and the overall decision space in SC2 is just a lot bigger.

I don't think this is true. Simple base building game like AoEs and Starcraft hardly need high level AI to do the maths on what buildings to build and in what order. It could be as simple as having a basic build queue as you reach each milestone. When I played AoKs it was pretty simple for a human player to get the most optimised balance as there were few factors to worry about.

Without a breakdown of both AIs and how they interact with the game space it would be hard to say which has to work harder. I'm pretty sure the battle mechanics of modern TW games are infinitely more complex and I assume that's why the used a more simple game to show it off.

Old AIs indeed could use a fairly straightforward scripted behavior for base building which is why they kinda suck (but it can be "good enough").

AlphaStar uses reinforcement learning, you can read about high level problem and approach here: https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

But you don't really need to know how they both work to tell that decision space in SC2 is just massively bigger. As they mention they have 10^26 legal moves at any point. Working with limited information and long term planning are also much less of a thing in TWW.

Oh, I think we almost certainly do need to know how each AI works and how useful are they for each game type. As the video mentions above, AlphaStar has serious limits and I've not been shown any evidence it would work for a more complex game like modern TW, I think there's a reason why they used it on a pretty basic RTS game.

Until it's actually tested on a more modern game like TW (not forgetting no one can replicate CA's programming - the reason they have no competition) I'm not going to accept it's some ultimate solve for TW's issues. If it's enough for you - cool.

TW is not a more complex game in terms of battle AI though, the only thing complex about it is the number of factions and units, but fundamentally it's a simpler.

But everyone's entitled to their opinion of course.

Is it? Any references? Here's the thoughts of a person with a doctorate in AI, who understands TW AI is cutting edge despite what the 'forum experts' say. Opinions are great, as long as they are backed up with something.

https://www.youtube.com/watch?v=XBzTJOYgW0M

Simple games like the above or AoE/K use simple AI because that's all is required for them.

As mentioned above, AlphaStar has clear limits as it only seems able to learn the most efficient way to win on one particular game and I've not seen it perform on any other. AlphaStar also isn't being used to make games 'better' or more enjoyable, it's there to test the quality of AI to do certain tasks, as in efficiency.

Would it make TW a better game? Dunno - more **evidence** required.

AlphaStar was a project made by the same team that made AlphaGo, which beats the best players in the world at Go, one of the most complex classical turn based strategy games due to huge decision space. If you understand how neural networks and reinforcement learning work, you could see how it can be applied to basically any game or a lot of things outside gaming as well. And it's not hard to see that the number of inputs and decision space in TWW battles is a lot smaller than in SC2. But to understand that, you need to understand how these things work so the only reference I can give you is this https://www.coursera.org/specializations/reinforcement-learning but it's a lot of time investment and you'll have to learn basic ML first. Other than that no, I don't have a video of Andrew NG saying that Starcraft is more complex than TWW because he probably doesn't even know about TWW.

As for making a better game - it's a separate question, players might not find playing against an AI that always outplays them enjoyable.

Games like Go are less complex than basic RTS games like AoKs and Starcraft 2, so let's not take the conversation backwards. Using the idea that the programme is more complex because it beat something less complex tells me you know less about the subject than you think.

Again, you've not shown any evidence that battles are more complex in the 11 year old game SC2 so I'm not going to reply on that until you do.

Again, you missed the point from the expert opinion above that machine learning can only follow simple rules like win in the most efficient way. It cannot be tweeked in a meaningful way so would be completely useless for game design.

Yes, you have no evidence that a basic RTS is more complex than an up to date TW game - that's my point and why I'm rather confused you seem certain about it. To give a reasonable argument I assume you would need to count up the unit types, the effects of positioning, aura abilities, constant abilities, possible combos, use of magic, charge, leadership, elevation, fatigue etc.

What has what and what doesn't and how complex is it? Questions like that around a coherent argument.

I used Go as an example to show that these approaches are applicable universally, not as a more complex game. Obviously they moved from less complex Go to more complex SC2.

And quoting my edit from the previous post "But tbh you don't even have to know anything about ML and just play both SC2 and TWW at a decent level to tell that TWW battles are not nearly as complex as SC2."

SC2 has (compared to TWW battles):
-strategic layer and long term planning
-scouting and working with incomplete information (to a much greater extent)
-just more **** going on with a lot more units, buildings and terrain features, and the number of possible actions you can take at any moment is just a lot higher

AlphaStar models are different models trained on a specific MU (e.g. Protoss vs Protoss), so the fact that TWW has more units and factions mostly means they would need more models, but it's just quantity, it's not fundamentally more difficult. And each model would be simpler due to reasons above.

Anyway, if you're really interested in the topic I suggest you reading up on the subject, I understand that not everyone wants to or has the time to learn ML but there's enough popular articles to at least get a high level understanding.

If you're here just to argue let's just agree to disagree.

Yes, a much less complex game which doesn't forward your point.

strategic layer and long term planning - The video above makes it clear that TW battles work on a tactical and strategic level. Strategic means an overall strategy which anyone playing a battle will use.

scouting and working with incomplete information (to a much greater extent) - You can't see what covered units are doing as well as stealth units. An AI either has to deal with it or doesn't, it can't see all the units for numerous reasons or what the player will do. An issue for both games.

just more **** going on with a lot more units, buildings and terrain features, and the number of possible actions you can take at any moment is just a lot higher - BS, WHTW has a lot more units with more complex and numerous abilities, magic, attack options etc.

The fact they have more models with differing abilities, thanks to the impact of the campaign map there can be random abilities placed on units, LLs etc. Of course more unit types add to the complexity.

You're the one being triggered because I don't think you know as much as you think you do on the subject. I love the way you completely ignore the issues brought up by the expert above.

I don't think this program can sort out the problems with TW, this is backed up by the expert above. As far as I'm concerned it's over.

Seswatha#7633

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?

The funny thing about AlphaStar is that the way it plays is kinda similar to either current or some of the older meta, though there are some noticeable differences.

And that only started to happen when they limited its micro, because at unlimited APM it would just own everything by spamming blink stalkers and microing them perfectly.

Some of the notable differences in SC2 are that it oversaturates bases with workers a little and uses static defenses a lot more.

Since TWW community is smaller and the number of MUs is bigger I'm pretty sure the difference with TWW meta would be a lot more staggering though.

WojmirVonCarstein

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.