AI tends to attack the same position (at all cost) over and over again.

[exsonic01]

Based on my experience in this game so far, I found that the AI (regardless of factions) likes to push to the same way, same road, or same positions inside a scenario.

 

Once I firstly take the "sweet spot" of the map (some important buildings, important crossroad, or important openings in the woods, treelines) faster than AI, the AI get ambushed and took heavy casualties by my troops. After then, usually real world counterparts (and humans) choose different positions to attack, circumvent to hit the flank, or try to become far more sneaky with the plan B. On the other hand, it seems that the game's AI push his 2nd / 3rd wave with the very same points with the same attack speed and very same way with the first wave, though the AI knows his troops were annihilated there, and my atgm / tanks are hiding there.

 

It is very hard to describe or judge everything of AI about attack, since AI only "tends to" choice the same plan with the first wave, which means it is not "always" happening. However, I think (or I feel) AI likes to stands with his original plan of the first wave, doing no precaution with my already-known-ambushed positions (only arty shells + air strikes are everything I can expect). No flanking or no luring or no sneaky back-stabbing. ( If it was multiplayer, I should move my positions before any flanking attempts. )

 

I know it is hard to make AI very smart like that, but after I watched 2 atgm teams + 1 tunguska in good spot became the hero of Russian federation by murdering tons of Strykers + US riflemans in the same ambush position, (It was "rolling on the river" battle and I was playing as Russians) I wish the AI in this game more sneakier, deadlier, and smarter then now. They just keep pushed the same position with quick moving speed without any other supports over and over again to my AT nest. I was in warrior difficulty. Is it difficult to make AI back-stab or flank the player's position under specific conditions?

[Apocal]

 

Is it difficult to make AI back-stab or flank the player's position under specific conditions?

 

Very difficult. The way AI plans work is all steps are run sequentially, with no branching decision tree. You can have a time delay or a terrain trigger (or both!) for the next step, but it is always 1,2,3...8, never 1, 2, IF Casualties taken -2A; IF No Casualties -2B.

[Alexey K]

Actually, there is no AI per se above TacAI level. Everything written in AI plan is linear actions sequence and have zero adaptability.

Real AI woud be generatin it's own AI plans during the game as situation changes.\

 

It would be difficult to deliver good AI for this game though.

[Stagler]

Very difficult. The way AI plans work is all steps are run sequentially, with no branching decision tree. You can have a time delay or a terrain trigger (or both!) for the next step, but it is always 1,2,3...8, never 1, 2, IF Casualties taken -2A; IF No Casualties -2B.

 

That may present another development avenue.

 

IF not in next zone by 01:00 game time THEN etc etc

 

Things already run on time hurdles.

[Alexey K]

Very difficult. The way AI plans work is all steps are run sequentially, with no branching decision tree. You can have a time delay or a terrain trigger (or both!) for the next step, but it is always 1,2,3...8, never 1, 2, IF Casualties taken -2A; IF No Casualties -2B.

 

Scenario-hardcoded decision tree approach is too limited to develop really flexible and adaptable AI behavior. It would be very hard to create and maintain for scenario designer as well.

Scenario designer should give goals and constraints but not script every step and branch of plan, but that approach required dynamic AI coded into game engine.

[womble]

Based on my experience in this game so far, I found that the AI (regardless of factions) likes to push to the same way, same road, or same positions inside a scenario.

You're not wrong. That is typical AI behaviour. The times you see it not do this, it's probably because some other "Order Group" has come into the fray, or the original contingent now has an order that doesn't mean it needs to go through your position.

 

 

 I was in warrior difficulty.

"Difficulty level" has no impact whatsoever on the quality of the AI's "decisons", so that's not an issue.

 

 

 Is it difficult to make AI back-stab or flank the player's position under specific conditions?

It's mostly impossible, without giving the AI enough troops to have effectively separate, fresh forces for each "specific condition" you want to handle. You can make it do some clever things with triggers, but, as has been said, every order group follows a single, linear order path, and all that Triggers can, to date, do is change the timings of when they take the next step on their path. I'm sure that branching AI plans are being worked on and will eventually materialise.

 

 

Actually, there is no AI per se above TacAI level. Everything written in AI plan is linear actions sequence and have zero adaptability.

 

While what you say about linear sequence and lack of adaptibility is true, AIUI, there is an "executive" level of AI which turns the straightforward instructions of the plan into precise movement orders (waypoints you'd see on the map if you could peek at the AI's "order phase"). That "entity" is supposed to be able to coordinate fire and movement, from what I've read on here, but I don't see a great deal of evidence of that.

 

It would be difficult to deliver good AI for this game though.

Yes, it would. It's a very complicated game, with a very great many degrees of freedom to eliminate.

 

That may present another development avenue.

 

IF not in next zone by 01:00 game time THEN etc etc

 

Things already run on time hurdles.

I would hope that's one of the early "branching conditions" they implement. "NOT" versions of the current time and trigger based "move to next step" seems like the most straightforward iteration.

[Alexey K]

 

Yes, it would. It's a very complicated game, with a very great many degrees of freedom to eliminate.

 

It can be crowdsourced to some extent. Consider having crowdsourced database of formalized METT-TC descriptions of tactical situations with local action plans.

Players can design and contribute their own plans. Alternatively, plans could be (semi)automaticly extracted from battle replays.

Depending on success of applications of that templates they are automaticly valued from worst to best. Worst are dropped over time to purge DB from rubbish plans.

 

P.S. Hey, I've just designed SkyNet!

[womble]

It can be crowdsourced to some extent. Consider having crowdsourced database of formalized METT-TC descriptions of tactical situations with local action plans.

Players can design and contribute their own plans. Alternatively, plans could be (semi)automaticly extracted from battle replays.

Depending on success of applications of that templates they are automaticly valued from worst to best. Worst are dropped over time to purge DB from rubbish plans.

 

P.S. Hey, I've just designed SkyNet!

Without better AI awareness, all such plans would be specific to the scenario, and pretty much couldn't address QB issues; that would make evolution slow. Having plans, and force composition as extractable elements that can be imported either wholesale or fragment-wise would be a step forward anyway. Along with a better orders UI, preferably built in the 3D environment.

[Alexey K]

Without better AI awareness, all such plans would be specific to the scenario, and pretty much couldn't address QB issues; that would make evolution slow. Having plans, and force composition as extractable elements that can be imported either wholesale or fragment-wise would be a step forward anyway. Along with a better orders UI, preferably built in the 3D environment.

 

Why specific? Their span should be limited. For example "taking small village" or "advance through open terrain".

These could be building blocks for higher level of strategy

Edited March 11, 2015 by Alexey K

[womble]

Why specific? Their span should be limited. For example "taking small village" or "advance through open terrain".

These could be building blocks for higher level of strategy

Because the variability in all those things is pretty huge, unless your AI can learn very general concepts, which requires very capable awareness of context. The senses have to be there to be able to detect the differences between "Small Village A" and "Small Village B". Or even "Small Village A from up a convex hill into a reverse slope defense" from "Small village A from up a concave hill where you can see and blast anything between you and the village, and the village itself. It's that sort of complexity which is always going to trip up AIs in CM.

[Alexey K]

Because the variability in all those things is pretty huge, unless your AI can learn very general concepts, which requires very capable awareness of context. The senses have to be there to be able to detect the differences between "Small Village A" and "Small Village B". Or even "Small Village A from up a convex hill into a reverse slope defense" from "Small village A from up a concave hill where you can see and blast anything between you and the village, and the village itself. It's that sort of complexity which is always going to trip up AIs in CM.

 

Real trick is to formalize METT-TC and actions plan into quantitive variables. After that engine can perform analysis on arrays of known rules (METT-TC -> Action Plan) and interpolate action plan for known situation. Another approach is to write AI with hardcoded knowlege of tactics, unit properties, etc. Both ways offer great challenge

 

EDIT: BTW, creating good AI can be fruitful in another way. It might remove need for micromanagement. Player could give platoons orders like "assault that village, defend this one"

Edited March 11, 2015 by Alexey K

[Baneman]

... It might remove need for micromanagement. Player could give platoons orders like "assault that village, defend this one"

 

For me, that wouldn't be CM anymore.

 

I want to be telling my men where to go ( maybe it's coming from a figures-wargaming background   )

[TheVulture]

Why specific? Their span should be limited. For example "taking small village" or "advance through open terrain".

These could be building blocks for higher level of strategy

 

I think the idea that there is a scripted set of actions that could be applied to "take a small village" or "advance through open terrain" without regard to the details of the map is entirely wrong. It require the whole gamut of synchronising in time and space the movements and positions of a whole range of units, and doing terrain analysis to determine fighting positions, cover positions, covered routes with respect to possible enemy positions etc.

 

Taking a small village isn't a limited task. If you can write an AI to do that, you've got 99.9% of the functioning AI needed for the whole game already. Or you've written something that is only going to work on very specific map configurations and hence, as womble says, you've got something where you need a specific plan per map in practice.

[JonS]

Real trick is to formalize METT-TC and actions plan into quantitative variables. After that [the] engine can ...

 Heh. That outline reminds me of this foolproof business plan:

 

Edited March 11, 2015 by JonS

[panzersaurkrautwerfer]

Concur on underpants gnome planning cycle. It isn't like there's a ready AI waiting to be dropped in.

[Alexey K]

 Heh. That outline reminds me of this foolproof business plan:

 

 

Yep, that was exactly my plan!

[Alexey K]

Concur on underpants gnome planning cycle. It isn't like there's a ready AI waiting to be dropped in.

 

Development of good AI would be very risky and expensive investment.

If I were Battlefront busines guy I would probably turn down such feature request

 

But here we can dream a littte, can't we?

[Apocal]

Scenario-hardcoded decision tree approach is too limited to develop really flexible and adaptable AI behavior. It would be very hard to create and maintain for scenario designer as well.

Scenario designer should give goals and constraints but not script every step and branch of plan, but that approach required dynamic AI coded into game engine.

 

A less flexible approach is already how scenario designers do things and I can't imagine a branching structure would be so ponderous to implement that it overrides the advantage not following a fixed step-by-step without accounting for the enemy at all.

Edited March 11, 2015 by Apocal

[Alexey K]

I think the idea that there is a scripted set of actions that could be applied to "take a small village" or "advance through open terrain" without regard to the details of the map is entirely wrong. It require the whole gamut of synchronising in time and space the movements and positions of a whole range of units, and doing terrain analysis to determine fighting positions, cover positions, covered routes with respect to possible enemy positions etc.

 

Taking a small village isn't a limited task. If you can write an AI to do that, you've got 99.9% of the functioning AI needed for the whole game already. Or you've written something that is only going to work on very specific map configurations and hence, as womble says, you've got something where you need a specific plan per map in practice.

 

You've missed my point. Idea is that AI is supplied with array of ready solutions provided by players and/or extracted from replays. By using machine learning algorithms it infers new solution to specific situation and puts it into action.

[Thewood1]

I think maybe supporting multiple cores and a 64 bit OS would come before "machine learning".  To me, there is some basic blocking and tackling that should be done first.

[dan/california]

You've missed my point. Idea is that AI is supplied with array of ready solutions provided by players and/or extracted from replays. By using machine learning algorithms it infers new solution to specific situation and puts it into action.

This has been discussed before quite a bit.  What you are asking for MAY be possible.  But it would require thousands and thousands of recorded games, many man years of programming, and probably more money for renting a supercomputer cluster than BFC has ever made.  I sure that this list is incomplete, but I am quite certain everything on it would be necessary.  The resultant AI code might or might not run on a home PC before 2040. Oh, and it might not work.  

 

This would really be a blue sky research project more than anything else.  I don't think BFCs fantasy Pentagon contract would cover it, in all honesty.  The only way it might, maybe work, resource wise, is too corral a bunch of compsci phd candidates into taking on the project.  That isn't nearly as simple as it sounds, probably still not cheap, and definitely not quick.

[Thewood1]

Let me give you an example of machine learning in use...

 

I have a GPS on my car.  It has a machine learning capability to offer a route and then after "x" repetitions of me driving my own route, it will "learn that route and offer it and variations to me in the future.  I have owned this car for two years and made hundreds of GPS-guided drives home.  It has yet to offer me the route home that is most efficient.  It has altered its offered route home slightly, but still takes me 5 miles out of the way even though it can see every road it needs to take on my route.  I have asked the dealer about it and the tech laughed at me and said it hasn't worked since it came on the market.

 

Machine leanring sounds good as an exercise, but applying it to a game is not going to be a trivial exercise.  That doesn't even take into account the expense and lost opportunity in doing it.