New CMx2 AI Improvements (planned)... New bone by Steve buried in PBEM noise thread

[aka_tom_w]

Battlefront.com

Administrator

Member # 42

posted March 01, 2005 11:45 PM                         

Juardis,

quote:

The question that BTS has yet to answer though, is there increased attention to the AI part of the game or is there really nothing more that can be done in that area?

Steve Says:

Oh, there is certainly more to be done in this area. Always is. I have mentioned some of them recently, but here are a couple:

1. Relative Spotting - makes the player less certain about what is going on, hinders his ability to be Borg/God, and in general causes him to be more in the dark (various options make this more or less true, player choice). Since higher level planning is the hardest AI behavior to program, reducing the player's ability to plan so well (i.e. know more than he should) automatically levels the playing field a bit more in favor of the AI.

2. More detailed C&C simulation - similar to above, the increased challenges that come with more realistic C&C mean that the Human player will find it more difficult to react to the AI in a Borg/God like way. This is also made more/less difficult through player choices, but since the players that are asking for a better challenge should be playing the game at its most challenging level, it is assumed they will be using such features (and if they don't, they have no grounds for complaints).

3. Scenario Editor Tools - there are a variety of things we can offer the Scenario Designers to coax better, more scenario specific behavior out of he AI. Currently there is no direct way to do this in CMx1, though there are certainly "dos and don'ts" the really experienced Scenario Designers have discovered through trial and error.

And lots of other stuff too, but I don't want to get into those things yet.

So... yes, the AI in CMx2 should be a lot more challenging than the AI in CMx1. Will it actually turn out that way? We won't know until it is programmed, but we'd be pretty bummed if it didn't.

Steve

[ March 02, 2005, 07:11 AM: Message edited by: aka_tom_w ]

[Thomm]

Points 1 and 2 sound more like dumbing down the human player than improving the AI.

Point 3 is promising.

All in all not enough substance here for specific discussion, IMHO.

I assume that SOPs will be implemented, which will automatically lead to better AI behavior at the platoon level.

Some form of learning would be great, although I doubt that I will play the scenarios more than once. I am thinking back to the CC4 campaign where I fought over the same map a dozen of times - had the AI remembered my deployment pattern, the Germans would have walked over me ...

Here is a good article for you:

http://www.gamasutra.com/features/20010912/sterren_01.htm

Best regards,

Thomm

[Thomm]

The same article, hopefully without registration:

www.cgf-ai.com/docs/gdc2001_paper.pdf

[Warren Peace]

I tend to agree with Rollstoy. The first two points really have to do with decreasing the human, not improving the AI. I also agree that the third point will help pre-made scenarios. However, to use Steve's terminology I would call point #3 a "cupholder" feature.

As I see it, we have to analyze what the AI currently does well and what it does not do well.

On the well side, the AI does use terrain well. (Infantry will always approach using maximum cover).

It does a good job of recognizing objective flags

On the negative side.

1) Does not use elevation

2) Poor coordination of attack (armor always goes first).

3) Does not maintain command lines.

4) Counter attacks too much when on defense (leaves fox-holes)

5) Poor use of Artillery.

6) Uses on-board mortars only in direct fire mode.

7) Has no concept of time (doesn't matter to AI the length of the scenario)

8) Has no concept of "fire-base" for heavy weapons

I know that making the AI better is difficult. However, I am dissapointed that none of these issues appear to be on the agenda

[Wartgamer]

I would be very interested in hearing some details about the C&C model they are improving for cmx2.

The present model in CMX1 is really just a squad/platoon model that has been blown up to multicompany/Battalion sized games. The fact that a player can gather information from each of his individual units, either from the 'God' movie or from individual units views and LOS during the orders phase, and then order individual units in any order he pleases and then even go back and edit those orders in light of further detailed info gathered from ordering orther units (delay time, LOS, etc). is mind boggling unrealistic.

There is no Company/Battalion HQ command decisions modeled or consequences. There is no Command FOW.

The product is a near simulation at small force levels and just a big game at large force levels.

I would like BFC to describe what they have in mind for C&C if they feel they are not giving away too much. I do not think that they are very receptive to ideas shot out in threads like these unless they describe where they want to go first.

[ March 02, 2005, 08:22 AM: Message edited by: Wartgamer ]

[Wartgamer]

And yes, 'dumbing' down the player means he is just too info smart. He is at a level of intel gathering not matched by any technology let alone WWII technology. He has a video network feeding him hyperinformation. Combine this with even normal human intelligence and the AI stands no chance.

An interesting thought is that if you make the game turns shorter, in the present CMX1 form, you just make the AI worse. Why? Because the superhuman gets to jump in and supermaximize all the uberinfo twice as fast. Want a better AI in the present game? Have yourself 'skip' a turn. Plan moves for 2 minutes length using the delay. Watch movie every turn but only give orders every other turn.

[aka_tom_w]

On the negative side.

0) Can we say it is too predictable?

1) Does not use elevation

2) Poor coordination of attack (armor always goes first).

3) Does not maintain command lines.

4) Counter attacks too much when on defense (leaves fox-holes)

5) Poor use of Artillery.

6) Uses on-board mortars only in direct fire mode.

7) Has no concept of time (doesn't matter to AI the length of the scenario)

8) Has no concept of "fire-base" for heavy weapons

I have NO idea how to program this but it lacks cunning, I would just like to be surprised by a really well co-ordinated counter attack!

I am hoping the Demo will feature a GREAT demo scenario akin to the likes of the CMBO Meeting Engagement Demo. (5 Sherms vs. 3 STuG's), where ALL that stuff Steve talked about with regard to "Scenario Editor Tools - there are a variety of things we can offer the Scenario Designers to coax better, more scenario specific behavior out of he AI", will all have been employed MOST wisely and shrewdly by the likes of RUNE and company, to knock a blow to the unsuspecting and even the most careful and skilled of us with a cunning and stunning AI defense and counter attack. Sure it will only surprise us once BUT I sure am looking forward to it.!!!

maybe?

-tom w

Originally posted by Warren Peace:

As I see it, we have to analyze what the AI currently does well and what it does not do well.

On the well side, the AI does use terrain well. (Infantry will always approach using maximum cover).

It does a good job of recognizing objective flags

On the negative side......

[ March 02, 2005, 09:14 AM: Message edited by: aka_tom_w ]

[Philippe]

Originally posted by Warren Peace:

7) Has no concept of time (doesn't matter to AI the length of the scenario)

I'm not so sure that's really a bad thing. The scenario time-limit is one of the most artificial and unrealistic elements of these battles, and what is needed is to give the player the same lack of a sense of time that the AI has. I would really love to see a scenario extension feature that worked a bit like the ceasefire or surrender button. While some of the problem can be placed on the doorstep of the scenario designer, very often realistic play is penalized by the time-limit, causing a scenario to come to an end in mid-firefight.

The battle should be over when it's over, not when the time limit does a lights-out number. That would normally mean that, for whatever reason, the attacker would stop attacking and pull back a bit, and the defender would either dig in or disengage. As it stands now there is not much incentive not to fight to the last man and the last bullet, simply because the penalties for thinking like an eighteenth century French monarch (Apres moi le deluge) are not severe enough.

Fixing this would make it possible to have scenarios about fighting retreats, withdrawals under fire, and, perhaps, scouting and reconnaissance.

[Wartgamer]

I believe that BFC said they want to get the AI to utilize more Borg/God info correctly.

This means to me that they just want the AI to abuse uberinfo just like the human. Hopefully, I am wrong because I do not like uberinfo to begin with.

One of the things I tried to impress upon the designer in the other thread was the concept of HQ based relative movie playbacks. The goal of the playback was a realistic situational report from the company commander's perspective. Its not just his POV but also all the info feed from his underlings in a slightly less detailed form than they are experiencing it (they see what they see during the order's phase but what they can convey to the company commander's movie may not be as 'fresh' or detailed).

The overall 'God' movie could still be displayed in a watered down form OR (for more realism) not seen as an option. So what is the effect of just seeing Company HQ relative movies? In some cases, you might have sections of your line 'go dark' (heavy attack/lost commo, etc). This means that you do not get to have a breaking news story at that section of the line. You got a problem. A realistic problem. Suddenly, you have Battalion commander headaches. The game has managed to force you to think like a Battalion commander for at least part of the turn.

Do you commit reserves? Create reserves? Send one of your staff there? Call for support?

Having a Battalion HQ 'movie' may not be a bad idea. The movie is actually dominated by sounds in some cases. I have read of many accounts where the sounds of grenades (they all sound about the same) was a good indicator of proximity to the enemy (either they are on your door step or you have got real close to them). High velocity guns are unmistakable, the intensity of the 'crackle' of mixed small arms indicates direction, etc. A Battalion Commander (or overall force commander) can be abstracted.

[ March 02, 2005, 07:51 AM: Message edited by: Wartgamer ]

[aka_tom_w]

Originally posted by Wartgamer:

I believe that BFC said they want to get the AI to utilize more Borg/God info correctly.

This means to me that they just want teh AI to abuse uberinfo just like the human. Hopefully, I am wrong because I do not like uberinfo to begin with.

I am not so sure that is what Steve said or meant. I have been following that discussion as well and if they could in fact hint or program the AI to use/abuse the God like player uberinfo I do not think that would be a bad thing, it should in fact make the AI somewhat more challenging and IMO more level the playing field between the "clever" human and the "clunky" AI. I don't think that is what they are intending to do however.

Steve said all players, at ALL times, under ALL settings will ALWAYS have to deal with the limits and restrictions of Relative Spotting and to be honest I think that GOES DOUBLE for the AI and I think that Relative Spotting code will hobble the AI worse than the clever player, BUT I TRULY hope I am dead WRONG about that

FWIW

-tom w

[ March 02, 2005, 07:48 AM: Message edited by: aka_tom_w ]

[Philippe]

An interesting feature of the old Battleground Napoleonic and ACW games was that you could turn control over most of your army to the TAC AI, limiting your role to, say, specifically that of a Corps commander while the larger battle raged around you.

Now I'll admit that I never really got the hang of playing that way, and from what I could tell the TAC AI did really stupid things. I don't think Tiller ported that one over with him to HPS, but the concept (though not the execution) was very good.

Personally I'm intrigued with the concept of figuring out how many new orders ( = how many ADC's with unblown horses do you have saddled up and ready to go at HQ) an HQ can really give, and at what distance. With WWII technology I suspect that the number is very low, simply because the typical company headquarters may have the odd radio, but apart from that is probably lacking in the mechanism of communication available to an eighteenth or early nineteenth century battlefield corps or division commander, even though the physical distances that need to be covered in both cases are about the same. If the master sergeant has laryngitis and the swift-legged corporal hasn't come back yet from delivering new orders to that machine gun up the road, it seems to me that units would either be acting on the basis of the micromanagement that you gave them at the begining of the game, or would surrender themselves to the control of the TAC AI.

This is very different from only making a move every two turns, because if you artificially eliminate one set of moves (which you probably shouldn't be allowed to make anyway, but that's a different story) the units that are not receiving orders will probably only behave in a convincing manner if they are in the middle of a route march across the map (not in column, of course, because CMx1 AI doesn't handle column movement very well) and aren't under fire. Steve and Charles don't need any help in coming up with a list of things that would trigger the TAC AI taking control. I just hope that they can get the TAC AI up to snuff to the point where command is by exception rather than the norm.

An odd, useless thought occurs to me as I write this. Supposing that one of the things that gets transmitted down the chain of command is a micromanagement opportunity. Normally a unit would be under the last set of detailed orders given (probably at scenario start) or the TAC AI. But the arrival of a radio message at platoon hq (or the arrival of a 1:1 swift-legged corporal, or perhaps some semaphore waving), could temporarily lift the micromangement ban on that hq and every unit under it that was still in command range (and I'm still not sure I understand why some command ranges are longer than others if it isn't defined by the strength of someone's vocal chords).

If you use an approach like that you'll be starting to model the behavior of units larger than a platoon. And the pace of battles will slow down a bit, simply because it will be harder to get your platoons to maneuver. You will also be starting to focus in on the difference between between a modern army amply supplied with radios (lots of micro-management opportunities) and the Soviet army on the morning after the purges (lots of TAC AI control).

[Wartgamer]

Steve said all players, at ALL times, under ALL settings will ALWAYS have to deal with the limits and restrictions of Relative Spotting and to be honest I think that GOES DOUBLE for the AI and I think that Relative Spotting code will hobble the AI worse than the clever player, BUT I TRULY hope I am dead WRONG about that

FWIW

-tom w

I don't want to rehash some arguments but I have my doubts.

How and If the AI can play under the most realistic settings will be interesting. Does the game 'hobble' the human enough under the most 'realistic' setting? Can the AI get a battlefield 'picture' in its 'head'?

Lets take the case of the No-God movie setting. How does the AI develop a concept of where the enemy is? He is staring at a blank battlefield except for his guys? Does he start cranking out a perception by jumping from unit to unit and 'building' a database from all those perspectives? Is he insane?

[Juardis]

Originally posted by aka_tom_w:

</font><blockquote>quote:</font><hr />Originally posted by Wartgamer:

I believe that BFC said they want to get the AI to utilize more Borg/God info correctly.

This means to me that they just want teh AI to abuse uberinfo just like the human. Hopefully, I am wrong because I do not like uberinfo to begin with.

I am not so sure that is what Steve said or meant. I have been following that discussion as well and if they could in fact hint or program the AI to use/abuse the God like player uberinfo I do not think that would be a bad thing, it should in fact make the AI somewhat more challenging and IMO more level the playing field between the "clever" human and the "clunky" AI. I don't think that is what they are intending to do however.

-tom w </font>

[aka_tom_w]

Originally posted by Juardis:

My concern, as stated in the PBEM in CMx2 thread, was that all the cpu cycles would be spent on LOS checks and overhead such that not enough attention would be spent on the AI (tactical and strategic).

From what I understand about how computers work I am not concerned about this possible issue at all. THERE never has been a time limit on the "crunch". If there is a Crunch cycle or phase in CMx2 then it will take as LONG AS IT TAKES and any LOS Relative spotting check will just get processed along with what ever cpu cycles the AI requires to get ITS job done. I do not, in any way, see this as an "either or" problem.

The Crunch will simply take longer and the faster computer you are using the better off you will be

This issue does not concern me in the least.

-tom w

[Wartgamer]

My concern, as stated in the PBEM in CMx2 thread, was that all the cpu cycles would be spent on LOS checks and overhead such that not enough attention would be spent on the AI (tactical and strategic).

I suppose that is very low level but you bring up a 'point'. A game, like CM 'WEGO', is unique from something like an aircraft control algorythm. In something like an aircraft, you have one second to crunch numbers and then put that 1 seconds worth of cpu cycles into actuating something.

In a game, one second of game time can be crunched in as long as it takes. Its open ended but the only consideration is the acceptable amount of time a customer can wait, the amount of programming/testing you can do and ingenuity.

An AI that can only play low realistic settings in some competant manner will drive the desire for more human opponents. That is, the only way to enjoy the realistic settings will be against humans. This will intensify a desire for TCP/IP if that is the only option.

A game like Close Combat was more 'aircraft' than CM. It had a time budget to get many things resolved in a time constrained fashion.

[ March 02, 2005, 09:55 AM: Message edited by: Wartgamer ]

[aka_tom_w]

I do not have a problem with allowing the AI to "cheat". If it makes it better, I'm all for it. For example, if the AI needs "Godlike" powers to see my weak spots then tries to exploit my weak spots, that surely would make for a challenging battle (and rewarding if I successfully pull it off).

-Juardis

That sounds good to me.

I am hoping there is a surprisingly large number of us that feel that way. (BUT I am not sure even that would help )

So far Steve as not commented at all about the possibility of assymetrical realism or FOW settings for the AI and the human player or two human players.

THAT one feature alone would be a break through of magnificient proportions!

-tom w

[Wartgamer]

If there's anything I remember from way back when was that BFC denounced other companies that had games that used cheating AI.

While having async FOW levels is not really a cheat, it may be needed so that the AI is not forced to be some insane info hound looking at everything it can. If it does that, it will, in effect just reduce its FOW level through intelligent examination anyway. So cut to the chase perhaps?

[Sergei]

You guys haven't been reading Steve's posts in the previous months well enough. In one of them he explicitly said that they consider letting scenario designers set AI special waypoints and giving players the option of letting the AI for example have partial information about the human player forces at start up.

[Battlefront.com]

Wartgamer said:

Combine this with even normal human intelligence and the AI stands no chance.

This is the crux of the problem. Nobody has made an AI that can think like a Human. Chess AIs are good, but they do not "think" in the way we're talking about in wargaming terms. A good chess AI is nothing more than a very good crunch clearly defined and extremely limited options until it finds the best pattern of them to use. Wargaming AI is, unfortunately, more like trying to simulate a CEO of company and everybody that works for him.

The upshot is that the more the player's actions are dictated by his thought process, the more difficult it is to create an AI that can do anything more than pose a small challenge. The opposite is also true since the less control the player has over the way the battle unfolds, the less the player's superior intellect comes into play. And that is realistic since in the real world no one person has the sort of control over his force that the Human player does in CM.

As for the Scenario Tools being "cup holders"... oh boy is that ever wrong One of the major problems for AI is coming up with a plan based on highly variable situations. CM has more variables than most HUMANS can handle, so to expect the AI to be able to grasp all of these things better than a Human is simply unreasonable. The solution is to allow the Scenario Designer to customize the AI to the scenario and therefore remove some of the need for the AI to "wing it". It is one of the only, and I mean only, ways to improve the AI's performance in a given scenario. Wishful thinkers need not respond

The topic of cheating is a very interesting one that we haven't discussed here since the old days of pre and post release CMBO. The current CMx1 AI does not cheat, and neither will the CMx2 AI. It is the wrong way to go about AI design. However, there is something which on the surface is "cheating" but when looked at philosophically it is more like "simulated intuition".

How does a Human decide what to do in a given situation? Does he simply look at the information in front of him and stop right there? Not a good player, that's for sure. Instead he imagines various possibilities and formulates predictions, which in turn he bases his plans on. The better the imagination and the more insightful the predictions, the better the plans are likely to be. (all else being equal). This sort of thinking is not something we can program into the AI. And I've said this before... if we could do this, we'd stop making wargames and instead make billions producing AI products that nobody else is currently capable of doing

Any ninny can see that this important part of "thinking" is critical to how well an AI can work in a chaotic, unpredictable, and highly variable virtual world. And since we can't recreate this thinking in code, the AI is inherently at a disadvantage so fundamental in nature that it is basically impossible to come up with an AI that will pose more than a passing challenge to the average gamer. That is, of course, if we don't work around this problem.

The work around is to allow the AI to "guess" like a Human would, but without having to write an AI that would win us a Nobel One way to do this is to allow the AI some amount of information that the Human doesn't have in the strictest sense, but DOES have through intuition, experience, or an understanding of the game system's limitations. For example, if you are playing a 1500 point game where the enemy is on the offensive... when you see a company of infantry you can, roughly, figure out how much armor the guy has. The AI, in current CMx1, can not. But if the AI were allowed a certain, highly controlled method of "peeking" at the Human's units... well, that would even things up a bit. The Human's guess based on no peaking will probably be more useful, but it is still something that can boost the AI's decision making capabilities substantially without certain cheating.

By certain cheating I mean if the AI knows you have an unarmored HQ at coordinates 32,45. This is information that the Human player could never have through intuition, since even a good guess is still a guess whereas 32,45 is no guess at all. Therefore, allowing the AI that sort of information is not good at all. It is unrealistic and based on other games which do this VERY frustrating to the Human player. So we aren't talking about this sort of thing.

Instead the AI might know that the player has an unarmored HQ somewhere in the area of 32,45 after having legitimately spotted the rifle companies. This is something that a Human player would likely deduce, then from there check out the terrain and say "hmmm... if I were the other guy, where would I set up my HQ?". The player might then choose a spot and hit it with artillery, or perhaps utilize some sort of recon asset to see if the guess has merit, then act if there is confirmation.

In the end there is only so much we can do. AIs, contrary to hardcore Grog thinking, don't sell games. They don't even hold back sales of games when the AI is discovered to be horrible (like Warcraft II, for instance). But we don't like to avoid trying things simply because everybody else and their brother does. So we'll do the best we can and for sure produce a better AI than other games. But will it be the wet dream of perfection Grogs demand? No. That is impossible and therefore we'd be fools to say we can do it.

Steve

[aka_tom_w]

Originally posted by Sergei:

You guys haven't been reading Steve's posts in the previous months well enough. In one of them he explicitly said that they consider letting scenario designers set AI special waypoints and giving players the option of letting the AI for example have partial information about the human player forces at start up.

I am surprised by this

I think I have been reading all the threads and all the bones/hints carefully but no where did I see any suggestion about how the AI might be set by the player to use a less realistic or more favourable FOW setting, different from the player's FOW setting.

Sergei, can you direct me to the thread or post where Steve said this: "giving players the option of letting the AI for example have partial information about the human player forces at start up. "

(for the first time other than his post below?)

Thanks

-tom w

[Sergei]

How would I find it, it was several months ago.

[Wartgamer]

Originally posted by Sergei:

You guys haven't been reading Steve's posts in the previous months well enough. In one of them he explicitly said that they consider letting scenario designers set AI special waypoints and giving players the option of letting the AI for example have partial information about the human player forces at start up.

I must have missed that also. But it does, just like tying my shoelaces and picking my nose, sparks me up another idea.

What if there were another new form of human-human play? What if the human on one side just outlines waypoints, suspected enmey positions, sequence of victory condition 'takedown', FO waypoints/release (he don't move till 'Flag' one gets taken down,e ct).

In other words, interject soem good old brains into your buddy's solo games? Even if the Human just gave the overall starting strategy, and since the games are not going to be that 'strategic', this may be enough to get some competitive play going. Another alternative is to have the Human jump back in after a victory location has changed hands and a new 'gameplan' needs to be interjected. You email your buddy and ask "Will you please tie my AI's shoes and blow his nose pleeez?"

If its left up to just the scenario designer, it may get stale and predicatable. I would personally like to see the interim movies and how my AI game plan panned out against the other guy.

[ March 02, 2005, 11:57 AM: Message edited by: Wartgamer ]

[Thomm]

Originally posted by Wartgamer:

If its left up to just the scenario designer, it may get stale and predicatable.

Please keep in mind that there are thousands of scenarios floating around, so I am not afraid that the need to play any CMX2 scenario twice or more often will become urgent! That is, unless you fail to beat that scenario in the first place, what should be a possibility given well laid-out AI pre-planning.

Best regards,

Thomm

[Wartgamer]

So there will be a scenario editor for cmx2?

You seemed to have missed the main idea of the post. A new form of Human-Human play.

[aka_tom_w]

Originally posted by Wartgamer:

So there will be a scenario editor for cmx2?

You seemed to have missed the main idea of the post. A new form of Human-Human play.

AND just to keep this discusion interesting (all the while off topic too I guess).

ONE other GREAT idea for Human to Human Play is by AI proxy ...where by I design and set up and an attack in the scenario editor and my opponent sets up the other forces in a defensive postion (in a way "programing" the AI in the editor) and we then let the AI handle both sides and sit back and watch the outcome of the battle. Hinted AI vs Hinted AI n

Crazy?

(oh Come on! it has a sort of Spy Vs Spy ring to it don't you think, YES that was Spy vs SPy from Mad Magazine he he)

OK its out there..

But some board game wargame geeks "might" think like this and if you let the game crunch ALL night (for instance) AND if the game had fully movie play back then you could wake up in the morning and see the WHOLE movie and see how the battle plan went?

It would be like a NEW form of TV

(OK ok sure...maybe that was ALL crazy but I do in fact think that was a form of lateral thinking or thinking out of the box that Steve was looking for in the GRAND scheme of things? :confused: maybe? )

-tom w