The AI in QB maps

[Freyberg]

Recently, in the FI forum, some of us were sharing our experience of programming the AI when designing QB maps.

This is an interesting topic that I think should be immortalised in its own thread

I've put my own meagre experimentation first (because it was first in that particular thread chronologically), but @kohlenklau has observed the behaviour of the AI much more scientifically; and I know there are others on the forum (including experienced designers like @benpark) who have also carefully observed how the AI allocates AI groups to forces.

I'm much lazier, I just playtest them, and if I have fun I give it a tick (also too lazy to search the forum for previous discussions of this topic, although I know it has cropped up in various contexts over years).

I'm really interested in what some of you more analytical designers have learnt, so I can apply it to my own maps

  

On 7/21/2021 at 10:55 PM, Freyberg said:

Thanks heaps for your kind encouragement.

As for the AI orders. I don't have the blueprint for how CM uses AI groups in QBs, but I've fiddled around a lot. I've heard the theory that the AI only chooses a small number of AI groups and ignores the rest, and I don't know for a fact that this theory is wrong, but I have found that giving the AI more groups leads to a more more satisfying game.

1) AI on the Attack (human player defending)

• I played each of these maps in at least one of their incarnations (Assault, Attack, Probe), and I based the AI attack roughly on my own attack strategy. I didn't test them all on the defence (AI attacking), but I did enough to get a feel for what would be fun (the AI is not at its best on attack, but some of them were pretty exciting).
• I never used less than 6 AI groups, typically at least 2 axes of attack (more with bigger maps), each with echelons, so that there is always overwatch, and the attacks are complex and based on AI groups going for key terrain, not just charging forward.
• I used a LOT of 'Maximum Assault' - counter-intuitively, I learnt in one of the threads that this command is more like 'Hunt' than 'Assault'; the AI will not banzai forward but instead will stop and engage.
• The attacks are relatively slow and measured (though hopefully not boringly slow) - they are designed to fit within the recommended time-frame of the map. There should be long sequences where the AI is creeping forward from cover to cover, pausing to fire while other elements move forward. I really tried hard to avoid Turkey Shoots.

2) AI on the Defence (human player attacking)

• In each plan, I always set up at least one AI group, covering the entire map, with no movement orders, set to 'ambush 1000m', in case the player has AT guns (which I personally love to allocate to AI defender). I was trying to do my best to avoid the 'gun limbered' situation (this still cropped up once or twice, but it was rare - I don't know how to eliminate it completely). I've found if I give the AI complete control of the map, (a) it will surprise me with gun placement (and I made these maps more for myself than anyone else, and (2) it does a good job of covering the approaches to the Objectives.
• In each plan, I usually set up one AI group, covering the whole map, with no movement orders, set to 'normal', in case the AI needed to place mines (as above, I wanted it to surprise me, so I could enjoy the map, and I think it does a good job).
• I never used 'Ambush Armour' at all, so the maps can be used for infantry-only battles.
• There are 4 defender AI plans for each map. One is usually quite static, sometimes with a little bit of fallback over time; one is usually aggressive, with spoiling attacks or counter attacks; the others are just what I felt like at the time, bits of fallback, bits of local counter-attacks, whatever.
• There are usually at least 6 AI groups (except on very small maps), or more. Bare bones defence is 2 static groups (see above), plus 2 axes of defence, each with 2 groups, so one should be covering the other's retreat or attack as they fall back or counter attack.
  (Again, I really can't tell how the AI allocates groups, but the more I played, the more I found the maps with more AI groups were more fun to play and more unpredictable, so despite being the designer I couldn't predict what the AI would do).
• Groups with movement orders usually have a larger number of short distance movements, executed reasonably frequently, rather than a small number of long movement orders (not always though, sometimes they'll flee to the back, regroup and counterattack; you never know). The reason for this was (a) to avoid the situation, particularly with armour, where the tank just sits there and lets you find its weak point - although there are some purely static defensive AI plans, most of them allow the enemy to reposition regularly, which hopefully makes it more fun; and (b) by using multiple, different AI groups, with overlapping movement and pauses, I wanted to avoid the Turkey Shoot counter-attack.

It doesn't always work, sometimes the AI redeploys somewhere dumb and gets itself killed - but it seemed to work well enough often enough that I enjoyed playing these maps myself, and despite having designed them, I generally didn't know what the AI was going to do next. Sometimes they were even really hard and I had to go back to a save-game to win

 

On 7/22/2021 at 12:58 AM, kohlenklau said:

Thanks for the report sir! 

I am now using all 16 AI slots so the fickle computer can pick and choose what it wants to use. I have a huge desert QB map and am running some tests. Taking notes like: Attacking AI used A1, A4, A6, A7 and A12! A1 was regt HQ Sherman and A Sqdn HQ. A4 was... etc. Previously I HAD a false assumption for QB's that A2 was always the support stuff like mortars. But I see that is not the case for AI on the attack.

If I see or don't see a selection pattern I will report back here for lack of a better place to discuss it. 

QB minimum time is 30 minutes so I THINK I see you design to that time scale and if player selects longer battle then everything after 30 is just the resolution of the struggle versus AI ordered/timed events?

 

10 hours ago, kohlenklau said:

I know what you mean. I think you did a great approach!

For my "North Africa" flavored QB with all 16 AI groups in my attacking plan, I ran another test with the computer selecting an all armor British attacking force. I was defending and watched in scenario author mode. The computer code of the QB picked what was nominally listed as an armored car squadron. 

A1 was 7 troop HQ. 3 vehicles.

A3 was the squadron HQ and 2 vehicles but also the 1 Troop HQ with 3 vehicles.

A4 was 5 troop HQ with 3 vehicles

A8 was 3 Troop HQ with 3 vehicles

A9 was 6 Troop HQ with 3 vehicles

A12 was 2 Troop HQ with 3 vehicles

The Game Engine 4 manual pdf says this:

Quick Battle Maps MUST have an AI plan for BOTH sides. You can use more than one plan and you can use as many groups in each plan as you like. The AI player will randomly assign units into groups.
Note: Since nobody knows which units will be taking part in a given Quick Battle, it makes sense to create AI plans in a much more generic way than for regular scenarios.

 

4 hours ago, kohlenklau said:

Another interesting observation...

For another huge North Afrika QB map of about 3000m x 2500mOn with battle set as huge and 1 hour in my QB set up settings, the defense AI observed in scenario mode used ALL 16 AI groups. IMHO, that is great news! That allows some cool little pockets of defensive action.

**************

@Freyberg I now return this thread to the regularly scheduled program... Thanks Frey!

 

 

[dbsapp]

I didn't get much, because I never edited QB (why would you do this?), but I appreciate the thread.

[kohlenklau]

1 hour ago, dbsapp said:

why would you do this?)

Good Question!

These are new maps that are then MADE into QB maps through various steps such as making AI plans, to offer out for use as Freyberg did for CMFI. I am trying to make some as well for North Afrika Desert Warfare.

This discussion of QB AI issues will be useful for the regular Joe Forumite if he has a map and wants to make it into a QB map -OR- he has a favorite QB map but maybe is disappointed with the AI's defense or attack...

 

[Freyberg]

3 hours ago, dbsapp said:

I didn't get much, because I never edited QB (why would you do this?), but I appreciate the thread.

Personally I find it fun - I started by adding terrain feature to existing QB maps I'd already played so I could play them as if they were new, and then just got more and more into it and enjoyed it...

2 hours ago, kohlenklau said:

This discussion of QB AI issues will be useful for the regular Joe Forumite if he has a map and wants to make it into a QB map -OR- he has a favorite QB map but maybe is disappointed with the AI's defense or attack...

...and the newer releases often have amazing Master Maps (the Fire and Rubble ones are very cool) - and it's fun to go through them, pick out some terrain you think would be interesting to fight over, put together a generic QB map (or several), leave it for a week until you've forgotten how you set it up (that step may require you to be a gamer-drinker, but I have that angle covered ), and you can choose exactly what sort of situations you want to game out, which is actually really fun.

[kohlenklau]

Observation: On my QB map set as huge in the design process, the size of the battle set by the player as he dials in his QB also dials back the number of AI groups used by the AI.

Huge: all 16 defensive AI groups used.

Medium: 13 used.

[Combatintman]

21 hours ago, Freyberg said:

Recently, in the FI forum, some of us were sharing our experience of programming the AI when designing QB maps.

This is an interesting topic that I think should be immortalised in its own thread

 

Totally agree mate - I think others have tried the same thing but they petered out.  I know the editor not quite inside out but not far off that; however, I have never really grasped QBs at all.  I'll be honest that this has partly been due to the fact that I rarely play QBs so have less interest in them than making scenarios but I think that I ought to understand the things a lot better because after putting the graft into making a decent map, it seems a waste not to be able to offer it as a QB map.

Good luck with the thread and I'm looking forward to learning from whatever contributions are offered by people who have walked the QB walk.

Have a like mate.

[Jace11]

@Freyberg

This is good info, I will check out the thread in detail. Thanks.

Not found a way to fix the game AI choosing 4 flammpanzer formations at 50 points each without ANY units in, by any chance?

In CMFB, I saw this today as a suggested formation. Why does it do that?

Edited August 16, 2021 by Jace11

[kohlenklau]

24 minutes ago, Jace11 said:

Why does it do that?

The interviewer should have asked THAT question!

[freetheEagle]

On 7/24/2021 at 2:50 AM, Freyberg said:

Recently, in the FI forum, some of us were sharing our experience of programming the AI when designing QB maps.

This is an interesting topic that I think should be immortalised in its own thread

I've put my own meagre experimentation first (because it was first in that particular thread chronologically), but @kohlenklau has observed the behaviour of the AI much more scientifically; and I know there are others on the forum (including experienced designers like @benpark) who have also carefully observed how the AI allocates AI groups to forces.

I'm much lazier, I just playtest them, and if I have fun I give it a tick (also too lazy to search the forum for previous discussions of this topic, although I know it has cropped up in various contexts over years).

I'm really interested in what some of you more analytical designers have learnt, so I can apply it to my own maps

  

 

 

 

 

 

interesting topic and yeah it deserves it's own thread 

[freetheEagle]

20 hours ago, freetheEagle said:

Recently, in the FI forum, some of us were sharing our experience of programming the AI Nox Vidmate VLC when designing QB maps.

This is an interesting topic that I think should be immortalised in its own thread

I've put my own meagre experimentation first (because it was first in that particular thread chronologically), but @kohlenklau has observed the behaviour of the AI much more scientifically; and I know there are others on the forum (including experienced designers like @benpark) who have also carefully observed how the AI allocates AI groups to forces.

I'm much lazier, I just playtest them, and if I have fun I give it a tick

 

20 hours ago, freetheEagle said:

interesting topic and yeah it deserves it's own thread 

I may be  able to help with it if you want