instead of AI what about mimic Intelligence?

[Smileyou]

we all know how hard AI is to script, build, make etc.. how about have every player send his play
to paradox server where such information would be transfered to "AI" that would "check" the
circumstances of your decisions and then basically mimic those decisions. of course just 1 guy
sending his 1 playthrough will mean nothing really, but if thousands sends thousands of playthroughs
a pattern can emerge, and just by mimicing those "decisions" that we make in certain decisions that
game AI can detect than by God we'll have maybe an unbeatable AI after all..

after certain amount of info to analyze paradox would send a patch to all the games to "update" the
mimic engine and voila.. in 4 months of intensive gaming i think if such "analyzer" of mimic-ish
software is possible than we can finally have a brutal machine capable of outdoing even the best players
and teams..

 

[Premu]

An interesting idea. Some points should be taken into consideration before trying to implement this:

1. Using machine learning techniques doesn't give you guarantee for the quality of the outcome. In some case the results are excellent, in others total failures. For example, the best AI for backgammon was created using machine learning. For chess on the other hand it simply didn't work until now.

2. You can't take normal single player saves for this purpose, as you very soon will run into a problem: The most successful games will be those that exploited known weaknesses of the AI. By altering the AI these strategies can end as complete failures. Also such an approach will fail for minor countries. For some countries like Albania there will be almost no games to analyze. Other minors who are chosen more often will suddenly become highly aggressive - as most players won't just wait until their faction leaders call for help. So in the end you will always have a huge Balkan war in 1938, killing the immersion for all players who want to stick to a more historical path.
Using multiplayer games could work, though - here you can only exploit game mechanisms but not the AI.

 

[Karlburg]

It's not that simple. Separating wheat from chaff in decisions is basically one of the hardest things for an AI to do and you'd be jacking up the amount of this by trying to tune it with input from players.

 

[jaava]

Recording entire games into some format and then mining anything meaningful out of that mess that is save-files... i wouldn't want to be the engineer for that job. man.. the files would have to huge to contain useful training data. You can do this in super mario, but i cant see it done in hoi.

 

[Atomcreator]

You would need a super computer. I guess China might lend you one.

 

[keynes2.0]

Yeah just sort this dataset by the "what the situation is" variable to divide it into categories and you know exactly when to mimic what behaviors.

 

[Smileyou]

its not like computer will need to analyse every bit of info, its easy for computer to "see" the gold from the crap if
you define what gold is.. gold in this situation would be (lets say a player plays as italy and sends his italy play
to the ai to analyse) when did the player(italy) send his troops to east to help Germany,what was circumstance of
italy then(supplies, resources etc, what danger to italy existed on borders(africa,UK,france etc), how many troops
in percentage to whole were sent to east?etc.. this are very easy for comp to find out and them implement same
strategy or same decisions in similar circumstance ..

as more and more playthroughs get analysed we do get more and more seemingly intelligent ai..

but yeah i agree, at some point it looks like a supercomputer would be needed..well, if that is what is needed
then dammit we need one) i think all PC Game developers should have a serious talk about this, this should be
a joint project..since all PC gaming and other gaming also relies on customers that are really not falling for the
ever dumber AI we get.. and supercomputer run by all game developers could be the answer.

1 Supercomputer could analyse all games, from all game developers with hundred of thousands of player inputs
in a second, .. so question is can they afford it? EA alone probably rakes in billions of clean profit and their games
hardly need AI but still i think would be desired to have as their game appeal would certainly be even bigger,not
to mention this much improved ai would open opportunity to make drastically different games(a game that is
truly open-aka open world but with computer people that would mimic real people means you can have a detective
story that you can actually investigate and interrogate people and really be in the wild as to who did it..)

so, 1 step, convince all pc developers and teams supercomputer for AI make happen is a must.

so question really is how much does supercomputer cost and how powerful does it need to be(like state of the art-
probably not since that would be a bit overkill plus would cost in the 5 bn $ range.. so a smaller one for 100-500 million$
should do it.. now, gaming industry is how much a year business? i heard somewhere its about 100 bn $.. while movie
industry is about 500 bn $.. meaning AI is a must, with movies we have real people acting, we can't have dimwits ai in
our games when gaming industry surpassed the movie one..

so 100 bn a year of profit means affording a ai analyse crunching super comp for 500 mill should be very affordable..

 

[Wizzington]

A learning AI is only as good as the learning algorithms you create for it. Using one for an entire game of any complexity is completely unfeasible just from the amount of work it'd take to create and refine the algorithms, not to speak of hardware and storage. I also doubt you'd even end up with anything useful in the end, even if you put in all that money and effort, because you'd up with this insanely complex thing that would have all sorts of learned bad behaviours simply by virtue of the fact that no learning algorithm can accurately measure success in a complex enough system. Even if you SOMEHOW achieved your goal and designed the perfect algorithms you'd end up with a horribly boring game because the AI would be virtually undefeatable.

In short: No. Mimicry is not intelligence.

 

[Smileyou]

aaa too bad, well i'm hopeful something big is bound to happen to AI department in lets say 5-10 years, simply by the power of logic of money interest
it must..as game industry overlaps in the future the movie industry it means tremendous money investment which will undoubtedly improve ai by bounds
and leaps..

in the end we may very well cry foul when ai will truly be unbeateable and will punish is bloodily for every single small mistake we do..and developers will
have to think how to "drowse" the ai so its not such a "nazi" all the time..

p.s.: yes, making algorithms for "learning" are very much difficult but mimicking ? this is totally different, its basically copy-cat of what people do,many players
do it in real life, we go through other people AAR and see what they build, when, how much, when to attack etc.. and then just repeat the process with a flavor of
ours here and there to keep it original.. see, even intelligent people copy..why should developers try to make ai learn anything, or even understand anything..it could be
the dumbest thing out there,without any concept of why to move 1 single unit, as long as mimicking tech software is state of the art i still see great future that is more
viable than making ai that "comprehends" stuff, and is logical let alone aware of itself..-skynet anyone?

 

[ringhloth]

The problem with your approach is that it requisites a program that can recognize good moves from bad in a way far more efficient than any program in existence. If you have that program, why not just have that program analyze the moves all by itself? Why do you need to funnel in permutations when you can just have the computer build them? Your approach also necessitates the saving of every move made in a game, which would snowball immensely and each save, which would likely need to happen somewhere between daily and weekly to not consume incredible amounts of RAM, would take 5-10 minutes, which is insufferable. One supercomputer cannot analyze every game because each game would need its own incredibly complex algorithm. Computers aren't nearly as good as humans at spotting patterns, and you'd essentially need to progress current models of computing 30-50 years.

Essentially, in order to do what you want to do, you'd have to have an actual AI, fully thinking. If PDS figures out how to build one of those, I don't think we'd see another video game from them. Johan will be too busy sipping mixed drinks on the beach of the private island he owns

 

[pheonicia]

Wouldn't then the goal be to create a set of algorithms just good enough to copy the general decision path and priorities of the player but not good enough to find/stick with the "perfect" strategy?

So, you know, mixing artificial intelligence with artificial stupidity to create sentience.
Though on second thought creating sentience from HoI's ai probably isn't a good idea.

 

[Darkath]

just type "selflearningai" in the console to enable it.

 

[SchwarzKatze]

Wouldn't then the goal be to create a set of algorithms just good enough to copy the general decision path and priorities of the player but not good enough to find/stick with the "perfect" strategy?

With our community and AARs here, the devs can actually see them instead of waiting for bot analysis. Also, it's not easy to tell what's good and what's bad in complex strategy games like this, especially around risky moves. Because a human could reload or even restart if the move fails, but if that happen in AI, you'll see totally opportunistic moves like those TW AIs. On the other hand, if the AI is too cautious and doesn't taks risks, you'll see things like Germany static in the Maginot or BoB like forever because the invasion windows human players found in AI France/UK don't happen if humans are playing them.

 

[Filou]

its not like computer will need to analyse every bit of info, its easy for computer to "see" the gold from the crap if
you define what gold is.. gold in this situation would be (lets say a player plays as italy and sends his italy play
to the ai to analyse) when did the player(italy) send his troops to east to help Germany,what was circumstance of
italy then(supplies, resources etc, what danger to italy existed on borders(africa,UK,france etc), how many troops
in percentage to whole were sent to east?etc.. this are very easy for comp to find out and them implement same
strategy or same decisions in similar circumstance ..

So then there is already an AI in place? Otherwise all the cpu players are just doing nothing. Which in turn means the player is doing moves in a vacuum, against a opponents that is 100% passive. They're not even building troops nor doing research.

So that play will be meaningless in terms of what happens outside of Italy.

But lets say the algorithm managed to learn how to run the basics of Italy with that one game (kudos on the programer btw!), is it enough for the rest of the countries to learn the basics of research and troop building? Hopefully it was enough for the UK, Germany and USSR to build armies and want to go at each-other, but somehow I doubt Germany will be much of a threat if they mimic Italian priorities.

So how many iteration do you expect before all countries have learned enough basics to give some minimal opposition? Because without it those moves from the Italy player are not much of a learning tool to the cpu.


As for your math about the money in the gaming industry... yeah... It's not like they're all sharing that pool of money.
Even if they were, the AI developed for Madden NFL might be shared with the one needed on FIFA, but I doubt it will be useful for strategy programers.

 

[Had a dad]

just type "selflearningai" in the console to enable it.

If I remember correctly, that is only a feature of EU IV

 

[Premu]

p.s.: yes, making algorithms for "learning" are very much difficult but mimicking ? this is totally different, its basically copy-cat of what people do,many players
do it in real life, we go through other people AAR and see what they build, when, how much, when to attack etc..

Mimicking is learning - you have to extract the actual strategy from the whole gameplay while also taking into account the circumstances in which the decisions were taken. Unlike a human who normally will also have the main strategy explained before actually applying it, plus additional knowledge about the game itself, a computer program will have to learn this stuff all on its own. In the end you'll need a really powerful learning algorithms, and those are not trivial, especially in such a complex game.

 

[Kovax]

I suppose that a computer such as what would be needed to handle this, as well as the self-learning AI for the job, has been available since 2001, when HAL9000 became sentient....at least in a movie.

I'd also hate to even think of the programming hours (and budget) which would be needed to write the software, in order to attempt to extract any semblance of meaningful information from the reams of data that would pile up over a few months of playing.


About the only "achievable" goal in that respect would be to have a cache of "previous games", where the AI could pick and choose from player-run country files, and choose one (hopefully with a better selection method than "random"), then follow that same set of moves for that country, regardless of how they meshed with the current player's moves for the player-run country, or with the other AI-run countries. So, Germany goes for a rapid "Seelow" instead of focusing on the East, while the SU launches an early pre-emptive strike on GER, and the UK decides to pursue a massive air buildup instead of a protective naval force, all taking wild risks because it's the most "interesting" thing to do, rather than the safest.

 

[Darkath]

A learning AI is only as good as the learning algorithms you create for it. Using one for an entire game of any complexity is completely unfeasible just from the amount of work it'd take to create and refine the algorithms, not to speak of hardware and storage. I also doubt you'd even end up with anything useful in the end, even if you put in all that money and effort, because you'd up with this insanely complex thing that would have all sorts of learned bad behaviours simply by virtue of the fact that no learning algorithm can accurately measure success in a complex enough system. Even if you SOMEHOW achieved your goal and designed the perfect algorithms you'd end up with a horribly boring game because the AI would be virtually undefeatable.

In short: No. Mimicry is not intelligence.

Isn't this technique used to an extent for facial recognition ? You feed a script thousands of pictures containing faces and thousands of picture not containing faces, and the sscript then is able to recognize faces on any picture.

This seems a very complex process for what just returns a boolean value (yes there is a face, or no there isn't). I think it's used widely in the domain of robotics to emulate sensorial feedback, but i don't see it applied to warfare simulation anytime soon.

 

[1alexey]

I suppose that a computer such as what would be needed to handle this, as well as the self-learning AI for the job, has been available since 2001, when HAL9000 became sentient....at least in a movie.

I'd also hate to even think of the programming hours (and budget) which would be needed to write the software, in order to attempt to extract any semblance of meaningful information from the reams of data that would pile up over a few months of playing.


About the only "achievable" goal in that respect would be to have a cache of "previous games", where the AI could pick and choose from player-run country files, and choose one (hopefully with a better selection method than "random"), then follow that same set of moves for that country, regardless of how they meshed with the current player's moves for the player-run country, or with the other AI-run countries. So, Germany goes for a rapid "Seelow" instead of focusing on the East, while the SU launches an early pre-emptive strike on GER, and the UK decides to pursue a massive air buildup instead of a protective naval force, all taking wild risks because it's the most "interesting" thing to do, rather than the safest.

Well, if the human-human "competitive" multiplayer games database is available, then at least AI could probably deduce both what is the likely continuation of current player unit build strategy, and what would be a better unit composition to beat it.

It is much more smaller thing, and AI would only need tables of unit of all sizes, and a score of how good did the game turned out for a particular player.

That is assuming AI knows how to handle basic unit types. Obviously a smaller taks, but much more doable, and would improve HOI AI difficulty by a huge lot.

 

[FOARP]

Basically, an AI good enough to distinguish the "good" moves from the "bad" ones would already be capable of beating the player.