The Ideal AI

[GermanPower]

Now before I start on my rant, here is somethings I want to point out.

-> Arguing over historical and ahistorical is ineffective and only makes the AI go further down the wrong direction, If you have objections to this go to one of the many threads and argue about it, not here please. (Sandbox versus Historical, same thing)

-> The world intelligent in AI should actual mean that

-> The ideas I'm about to present are always subject to modification

-> This could be for a HOI5 or a mod for HOI4, if it would ever exist as I fear HOI4 is already well on its way with what its doing, could 100% way totally wrong.

-> Devs only take in a fair bit of individual community input as at the end of the day its about making a good game for everyone and not just me. So what I'm looking for is people to add onto this idea and support it so that Devs may give it some thought. Most likely no one will care what I think but I'm taking a shot anyways ;D

-> Criticism is welcomed but blasted arguing and banter for the sake of arguing isn't welcomed. Please respect that.

-> The Key Idea behind what I'm about to say is about making the world feel real with the AI interactions

-> Game mechanics should almost always work in conjunction with making the world feel real and being immersed in it.

THE IDEA FOR AI, DUN DUN DUN!​

Let me cut straight to the chase, trying to confine the AI into the rigors of Sandbox (the madness of random) and the Historical (The bore of structure.) is only going to lead to less immerse AI and poorer quality, now I am not sure what they are planning for HOI4 but it shouldn't be done to the sandbox extreme or the historical end. Both make the game feel lifeless. Now what I'm about to say is somewhat sandboxish however it isn't at all what it is. Its about making the player feel like they are in World War 2, making the decision, keying the front, reacting to the AI!

Pathing and variability, the world you are playing in should react to you not just have a structured feel, if you do everything the way you things went historically things should turn out similar however if you start moving from what happens historical, which will happen eventually this should set off consequences and chain reactions within the game and cause the nations of the world to react to what you did and change the outcome of events and decisions by other nations, if enough steps are taken things might change history entirely. Your nation should be what shapes history and changes it when playing single player. You can always be historical if need be but by doing new things and different things you can warp and change how history plays out and alter what actually happens which forces new decisions on yourself and on the AI. There should be plethora of non history based decisions due to your meddling with history and events that should key off do to what you do and you should have many chances to change how things happen and be give options during events like the Fall of France and Poland, some mods already do this to a degree but it isn't done on a grand scale of pathing events and requirements for things to happen. For example the Veteran Army, Controlling of Ports, all of those effects should have real world effects and decisions as the world is shaping and changing during the war. This should also be in effect during peace time and the Ai should be reacting and to you and then you should be changing what you do due to the AI meddling. Lastly, things you do and things that happen should be meaningful and have great impact on the game and whats happening and change the decision making of the AI. This is my general idea of how AI should be done, feel free to post your thoughts, ideas or criticism.

Side note(More for the actual war): It always seems to me like I'm just playing through the same game with HOI3 and its not because the game doesn't throw random stuff at me, its in fact quite the opposite. Its the fact the AI simply doesn't do sensible things. They aren't ones to react to situations in a smart ways, the AI should force you to do things, just like you force them, it should be a relationship in which you are both reacting to each other over a grand scale. Now this is more for when you are already fighting while prewar should focus on diplomatic intrigue .

EX: If I shift to building thousands of Planes, the AI might build AA due and then we could go back in forth trying to match each other in various ways I might try and kill Convoys and force them to make other things like Great Britain and the glorious wooden plane or the Sten gun. There should be constant battling.

 

[jju_57]

One question. If the Ai is to respond to situations on the map then how do you handle fog of war and limited knowledge?

Also, you mix strategic AI and tactical AI without really getting into AI.

Finally, your post is about the player. How would your AI work if I'm playing the US? So much already happened by December 1941 so how does your AI deal with that?

EDIT: Actually that was more like three questions and an observation.

 

[Dalwin]

One question. If the Ai is to respond to situations on the map then how do you handle fog of war and limited knowledge?

Also, you mix strategic AI and tactical AI without really getting into AI.

Finally, your post is about the player. How would your AI work if I'm playing the US? So much already happened by December 1941 so how does your AI deal with that?

You mention a key point, the fact that there is both a strategic and tactical AI. This is often overlooked. The AI is not a person, it can be compartmentalized. Just because one part of the AI has access to information to make some mechanic work out well does not mean that all aspects of the AI have access to that same info.

If the game works better by having the strategic AI know how many tanks or carriers are being built by the opposition so that it can respond instead of blindly following some scripted path, I am all for that. Strat AI having an exact count of my tanks is not the same as the tactical AI having any idea where those tanks are. They are completely separate concepts. You do not even have to rationalize it.

 

[Pugmak]

One question. If the Ai is to respond to situations on the map then how do you handle fog of war and limited knowledge?

Also, you mix strategic AI and tactical AI without really getting into AI.

Finally, your post is about the player. How would your AI work if I'm playing the US? So much already happened by December 1941 so how does your AI deal with that?

EDIT: Actually that was more like three questions and an observation.

Here's how I think I'd try to address the AI, at least in ref to theater level, tacs and strat. Don't laugh. I'm not a coder or any such and really don't have a clue, but... here it is anyway.

1. Toggles, as follows:
A. Line of Battle. Means basically that. A toggle, set by the player, that designates an Army level HQ to form a line of battle and contend with what's in front of it, offense or defense as sub toggles. But, the Army AI has a set limit on spread (like areas of responsibility set for actual irl military formations, that takes into account its strength/size in divisions). Line of Battle units do not get yanked around to respond to threats outside their AOR.
B. Blitz. Offensive only. No regard for defending rear areas. Pure aggressive assault and penetration. Blitz units are also exempt from getting yanked around by events in theater other than their own AOR and target set by the player.
B1(because I just thought of this one and didn't want to re-tag the rest). Follow on. Force(s) designated by the player to follow the Blitz and secure flanks, etc.
C. Reserve. If at Army level, a Corps is held in reserve to play either line backer in defense, exploiter in offense, or general roust about in that Army's AOR. If at Army Group level, same, but with an Army.
D. Theater Reserve. Set at Theater level. A force, decided upon by the player, designated to respond to whatever weirdness the opfors cause to happen in Theater AOR.
E. Security. A force, set at Theater level to respond to issues affecting overall security of operations within that Theater. Rebels, safe guarding provinces of importance, protecting supply lines, etc.

Imo, the major problem currently with this, and all other game AIs, is that they're forced to contend with too much with too little direction from the player. I think that breaking it down into more manageable chunks would be a help for the AI. I also believe this might help the AI AI (non player) in giving better framework upon which the AI AI (non player) can build its own forces with an 'eye' toward the various functions, without having the entire force be, more or less, responsible for everything, all the time.

 

[Dalwin]

You do bring up some points.

I think one of the keys to creating a successful AI is compartmentalization. Different routines are responsible for different specific things. I am also fairly sure that this is what they are already doing. There is a routine that does nothing but decide what to build. It gets some input from what is going on instead of blindly following a preset script.

Along these lines of creating AI (talking true AI now and not merely an automated opponent in a game), the guy to watch is Ray Kurzwell (sp?). He is currently working on a project backed by Google to develop a real AI. Ray wrote a very interesting paper when he was at MIT in which he broke the human thought process down into many pattern recognition loops. Individual pattern recognition loops are relatively easy to program. Putting together a sufficiently complex web of them to achieve intelligence is theoretically possible.

 

[GermanPower]

One question. If the Ai is to respond to situations on the map then how do you handle fog of war and limited knowledge?

Also, you mix strategic AI and tactical AI without really getting into AI.

Finally, your post is about the player. How would your AI work if I'm playing the US? So much already happened by December 1941 so how does your AI deal with that?

EDIT: Actually that was more like three questions and an observation.

Interesting questions(please ask more!), Ummm here my off hand answers, there is already a intell system and this will effect how effective the AI is toward preparing and knowing what to do, this has already been hinted at in the actual game, however as it isn't near finished I'm not sure how far the rabbit hole goes with that. Imagine the AI gets bits of information on you and it then responds by building and preparing the appropriate units for its operations. We all know that Germany had horrible intelligence while the Allies had far superior. Many factors would influence it and for me it would require a robust and useful intell system, which might be happening already.

The US will have a variety of events happening to it based on whats going on, if Germany actually invades Poland, if France has fallen, and various factors, if Japan has actually invaded China. If Japan already won the war by 41, I don't think a Peral Harbor attack would happen, or if Japan had a more aggressive commander during Pearl Harbor due to the assignments of generals by players or AI if the history turned out differently,I (as Japan)would in vision a setup where you actually have to commit ships to commit pearl harbor and you could hand pick who is carrying out the operation and if you have a heavily aggressive general, he might have gone for that 3rd raid and be successful or they might have invaded, so much can happen.

Now thats the side for Japan, let me try and explain the side for the US, you would have a large group of options such as listening to a certain General whos screaming his head off to get the ships outta Pearl Harbor or if Great Britain has already fallen the US would have the option of establishing relations with Germany and keeping a Cold War Esq peace. Now as the US the reason certain things could happen if you chose to not support the Allies and keep more neutral, the situation really depends on what you did and if you altered from History and how much, each game will be vastly different due to how players are and this is why I like my idea for a AI and each time it will feel like a different game. Unless you pick a nation that has no real impact on the war....then things might simulate similar to the real history or if you attempt to be historical with a major nation, however your tactical capability of using units will further effect how things turn out in the world (What you research all that mess) and do to the many factors history might change and flip as you do things both tactically, Strategically and Diplomatically. If you give me a idea of what you would do for 5 years from 36-41 with the US I might be able to give you idea of how a world could pan out.

 

[Concept2D]

A few points

- Paradox's current AI currently do most of what you said.

- You do NOT want very good AI. Chess has very good AI, AI that can smash every single human on the planet close to 100% of the time, is that fun to play against ?.
-- Paradox needs to aim for fun for the majority of people at an acceptable cost.
--- Paradox has to try and make the Player feel their actions had a big effect, and that they are smart. If theres going to be a dumb party after a few hours of play it needs to be the AI and not the player.

- Sandbox and Historical is how Paradox decides to balance the scripts. For example Czechoslovakia has I think a 10% chance to reject the German offers in HOI3. Increase these percentages for more sandbox, reduce them for more historical. You seem to favour more sandbox as do I, but note a lot of people will rightly complain if not enough games have a Barbarossa etc.

- On the Planes/AA example I think this should be the next improvement area for Paradoxes AI. The HOI3 theatre system might actually do a version of this (I've done no research on it).
- Note it is a reasonably tricky problem.
-- Requires the AI to keep track of their enemies air strenght. The AI could cheat of course (by removing fog of war) but doing it properly requires
--- Keeping track of what your spies see
--- What your radar/troops see
--- making reasonable extrapolations on the info it has
-- It then needs to change it's change its build order to compensate, a simple way to do this is increase "build_anti_air" priority
--- A more effective one would be
---- The enemy has heavy armour researched 1 year ahead of time
----- Solution: Change Anti Armour and Heavy armour research priority to "2yr_ahead_of_time". Calculate when you catch up. Build Anti armour and HARM at this point, have practicals as high as reasonable. Would building CAS/TAC work?. Would blockading work ?, Would allies help ?, Would pressuring in the jungle help ?, Would it be more effective to just take the huge losses and continue with the current path? .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
- Do all the above for a reasonable cost both in development cost and CPU drain, all while making new and experienced player feel like superstars.

 

[Beagá]

You can´t compare doing AI for chess with AI for HOI. For starters it´s a real time game and timing (that is, the order which you send units) has Always been na Achilles heel for such games, regardless if it´s Starcraft or this.

I DO want na AI that finally does fighter sweeps and uses bombers properly instead of the idiotic german AI wasting the Luftwaffe pointlessly, or sending all the Kriegsmarine to the bottom in 1 month.

 

[Concept2D]

You can´t compare doing AI for chess with AI for HOI. For starters it´s a real time game and timing has Always been na Achilles heel for such games, regardless if it´s Starcraft or this.

I DO want na AI that finally does fighter sweeps and uses bombers properly instead of the idiotic german AI wasting the Luftwaffe pointlessly, or sending all the Kriegsmarine to the bottom in 1 month.

First chess is as real time as HOI, the clock is there for a reason.

Second the Starcraft AI is held back by Blizzard to increase the games popularity, for example compare the World Champ at the time to a simple mod

[video=youtube;DXUOWXidcY0]https://www.youtube.com/watch?v=DXUOWXidcY0[/video]

And before you say it would not be able to come up with build orders, the 7 roach rush build order was created by one of the genetic algorithm AI's.

 

[tommylotto]

I think you need to take into consideration what the developers are up against and what tools they have at their disposal. They could set up certain general rules of universal applicability such as x threat plus y balance of forces equals DOW. However that will result in sandbox behavior and very non-WW2 results. Or they can write scripts or event chains to keep the game on historical rails. If they do that, then the quality of the AI depends on the number of "what if's" they script for ahistorical actions taken by the player. What if an Italian player invades Yugoslavia in 1938? Will Bulgaria intervene? On which side? Should the decision be a fixed reaction or a percentage chance with modifiers to give the game some unexpected randomness? If random, what modifiers should effect the chances? When you think of all the what if's that would need to be accounted for, you can see how daunting a task that would be. So, an AI that allows for an historic outcome unless the player intervenes to change history is possible, just a mind numbing amount of work.

 

[Beagá]

First chess is as real time as HOI, the clock is there for a reason.



And before you say it would not be able to come up with build orders, the 7 roach rush build order was created by one of the genetic algorithm AI's.

Lol no. Chess is not real time. Units don´t move simultaneously, period. And thus timing is NOT an issue, or ath the very least is a different "type" of timing. So if I put Monopoly with a clock it becomes real time too? No comments.

Maybe I should use a better example then: Company of Heroes. Since Starcraft, specially 2 is about number crunchings and Actions per minute MUCH more than COH.

 

[GermanPower]

I think you need to take into consideration what the developers are up against and what tools they have at their disposal. They could set up certain general rules of universal applicability such as x threat plus y balance of forces equals DOW. However that will result in sandbox behavior and very non-WW2 results. Or they can write scripts or event chains to keep the game on historical rails. If they do that, then the quality of the AI depends on the number of "what if's" they script for ahistorical actions taken by the player. What if an Italian player invades Yugoslavia in 1938? Will Bulgaria intervene? On which side? Should the decision be a fixed reaction or a percentage chance with modifiers to give the game some unexpected randomness? If random, what modifiers should effect the chances? When you think of all the what if's that would need to be accounted for, you can see how daunting a task that would be. So, an AI that allows for an historic outcome unless the player intervenes to change history is possible, just a mind numbing amount of work.

Doing anything in a video game takes loads of work, not sure what your point is.

Concept what in the world are you talking about dude? No one has claimed make the AI god like, I've said realistic. Also I think you've missed the entire point of what I'm saying, its not sandbox at all what I'm saying, it may seem like it but its based around the decisions of a player. Paradox is doing this stuff? I'd like some links. I think you are trying to apply the sandbox label to the idea, its not that at all man.

 

[Dalwin]

Realtime is not the right term for differentiating chess from HOI, neither game is realtime in any sense. HOI just has a very high number of turns each representing an hour.

A better analogy would be to compare a single core cpu (chess) to a multithreaded one (HOI). There are many more things going on each turn in HOI than there are in chess, by two or three orders of magnitude. However, HOI is much more forgiving of minor mistakes than is chess. One badly placed pawn at high level play can easily mean you are heading toward a losing end game 30 moves down the road.

 

[tommylotto]

Doing anything in a video game takes loads of work, not sure what your point is.

Concept what in the world are you talking about dude? No one has claimed make the AI god like, I've said realistic. Also I think you've missed the entire point of what I'm saying, its not sandbox at all what I'm saying, it may seem like it but its based around the decisions of a player. Paradox is doing this stuff? I'd like some links. I think you are trying to apply the sandbox label to the idea, its not that at all man.

Pathing and variability, the world you are playing in should react to you not just have a structured feel, if you do everything the way you things went historically things should turn out similar however if you start moving from what happens historical, which will happen eventually this should set off consequences and chain reactions within the game and cause the nations of the world to react to what you did and change the outcome of events and decisions by other nations, if enough steps are taken things might change history entirely.

I think I know exactly what you are trying to get at, because that was what I tried to do with my mod. However, I only tried to do it from Italy's perspective. There are event chains for Italy intervening in Spain if the Republicans win and reinstalling Franco. There are event chains for the Balkan nations to gang up on Italy if Italy attempts to attack the Balkans in 1938. There are chains of events for Italy to intervene in the Anschluss in Austria. There are event chains for England intervening in Greece. There are event chains for France intervening in Yugoslavia. There are event chains for Stalin to get involved if Italy tries to conquer Bulgaria or Romania. There are event chains for Ethiopia. There are event chains if Italy is successful in acquiring territory from France before the fall. There are many many more. Each chain is a ton or work, because you need to account for as many variable as you can imagine and play test to see if there are absurd results. It requires a massive amount of man hours and like I said, that was just for Italy. You would have to do the same thing for each major, at least. You would need to brain storm what if's for each major and follow all those what if's to all of their hypothetical conclusions. Then you would need to write scripts or event chains for every single conceivable what if path. I would love it if Paradox were to go to that much trouble. They certainly know better what they are doing than I. I just think it is unrealistic to expect that they will.

 

[GermanPower]

I think I know exactly what you are trying to get at, because that was what I tried to do with my mod. However, I only tried to do it from Italy's perspective. There are event chains for Italy intervening in Spain if the Republicans win and reinstalling Franco. There are event chains for the Balkan nations to gang up on Italy if Italy attempts to attack the Balkans in 1938. There are chains of events for Italy to intervene in the Anschluss in Austria. There are event chains for England intervening in Greece. There are event chains for France intervening in Yugoslavia. There are event chains for Stalin to get involved if Italy tries to conquer Bulgaria or Romania. There are event chains for Ethiopia. There are event chains if Italy is successful in acquiring territory from France before the fall. There are many many more. Each chain is a ton or work, because you need to account for as many variable as you can imagine and play test to see if there are absurd results. It requires a massive amount of man hours and like I said, that was just for Italy. You would have to do the same thing for each major, at least. You would need to brain storm what if's for each major and follow all those what if's to all of their hypothetical conclusions. Then you would need to write scripts or event chains for every single conceivable what if path. I would love it if Paradox were to go to that much trouble. They certainly know better what they are doing than I. I just think it is unrealistic to expect that they will.

Concept what in the world are you talking about dude? No one has claimed make the AI god like, I've said realistic. Also I think you've missed the entire point of what I'm saying, its not sandbox at all what I'm saying, it may seem like it but its based around the decisions of a player. Paradox is doing this stuff? I'd like some links. I think you are trying to apply the sandbox label to the idea, its not that at all man. <------This wasn't for you, it was for Concept.

Again I haven't stated that I expect them to do anything, not sure where that came from. I specifically state that they aren't, this is just a idea to see who supports it. We are both in the same boat.

 

[Dalwin]

I am a little disappointed that there were no comebacks on my comment concerning Kurzwell's approach to developing AI. Even though he is working on developing true AI, the same approach, on a simplified level, involving a series of pattern recognition loops, could be applied to developing an automated game opponent. In fact, if you want your AI to be adaptive, to respond to your actions, I think this is the only valid approach.

Any other simply gives it a finite and relatively short list of preset reactions, most often triggered by randomness. Once you take this latter type of AI out of its comfort zone, get it away from its scripted responses, it flounders and becomes unfocused.

Here is a simple example of a pattern recognition loop. I see that I am taking significant losses to enemy subs. I respond by building more destroyers and researching ASW.

 

[jju_57]

Here is a simple example of a pattern recognition loop. I see that I am taking significant losses to enemy subs. I respond by building more destroyers and researching ASW.

Except that alternative of building more DD's has to be programmed in as an option. I can instead change routes, cancel the convoys, capture the bases the subs come from and maybe some other things.

Point is all alternatives still have to be programmed in, and evaluation or even perception has to be applied and then a path of action picked with future reevaluations.

I've been involved with computers and programming for close to 40 years now. Maybe I'm a pessimist but I don't see real AI anytime soon (not for the next 100 years or so if then). The level of processing will improve but not real AI as based on how humans evaluate and make decisions.

Think of this one simple item. A new idea. If the computer is restricted to just the information it has access to there is no possibility for new ideas or thoughts. Yet this happened in human nature constantly. Can a computer AI compose a song? I can program in all notes and also every song ever composed but there will not be a new song coming out. When that happens that is real AI.

 

[Dalwin]

Except that alternative of building more DD's has to be programmed in as an option. I can instead change routes, cancel the convoys, capture the bases the subs come from and maybe some other things.

Point is all alternatives still have to be programmed in, and evaluation or even perception has to be applied and then a path of action picked with future reevaluations.

I've been involved with computers and programming for close to 40 years now. Maybe I'm a pessimist but I don't see real AI anytime soon (not for the next 100 years or so if then). The level of processing will improve but not real AI as based on how humans evaluate and make decisions.

Think of this one simple item. A new idea. If the computer is restricted to just the information it has access to there is no possibility for new ideas or thoughts. Yet this happened in human nature constantly. Can a computer AI compose a song? I can program in all notes and also every song ever composed but there will not be a new song coming out. When that happens that is real AI.

Of course you can add refinements to my simple example. It was obviously labeled as an simplified example.

By adding other responses besides the two I listed you are doing only that, refining the pattern recognition loop. Your response does not really address the underlying concept.

At the end there you are jumping back to the idea of true AI vs an automated gaming opponent. This is mostly my fault since I start by making the leap between those two.

The process you are referring to for creating new things is accomplished by mixing parts of concepts from known things. This process is called synthesis and it is not impossible to program synthesis into a computer (though HOI needs nothing so fancy).

Also, by the way, computers have been composing original music for well over a decade. Taste is subjective, but I think some of it is better than what some human composers come up with. It takes more than one specialized area of creation to count as AI, though.

 

[GermanPower]

Well the point is really to just make the AI more responsive towards the individual. Its entirely possible not sure why you guys are getting into a sentient being. I'm keeping this within the confinds of whats possible and doable. What you guys are talking about is starting to leap outta bounds of the discussion, care to give me some examples of how history could be altered by the player or events the player might be able to manipulate?

 

[jju_57]

I've heard the AI composed music. I know taste is subjective but not a single one of these 'songs" has ever gotten beyond curiosity stage. You can do better by throwing a dart at random notes. There is no "art" (highest concept) in what the computers have done.

But back to point. Even an automated response based on patterns is too much code for a game like HOI3.

The chess vs. HOI examples were way off as chess has strict rules and a I go you go style. You can force calculate millions of paths and as long as your evaluation logic is good the computer has it easy. But it is NOT AI in any sense of the word. It is just more and better processing power to evaluate specific outcomes. The permutations might be large but not the complexity of the programming.

Programming a HOI AI requires complex programming logic.