Just how "I" is the AI?

[Cymru]

In a recent battle my armor was in trouble, so I ducked behind a building and hoped for a flank attack as he came by: which worked pretty well. Now; if I see an enemy tank disappear behind a building, I make a note and am a bit more careful when operating in that area.

So, does the AI 'remember' that tank, or is it strictly 'out of sight, out of mind'?

[womble]

ISTR that there is a very short term memory of what a contact was. But for the most part it's just a fading "?" once the asset is unspotted. This impression is gained from reading posts by people who seem like they orter know

Part of the AI's problem is that it will continue to follow its AI plan, even if the situation makes that plan somewhat risky, and it can't adapt. I had a situation where my infantry ambushed 3 StuGs. The first went down to a close assault. The second to a zook's third shot and the third backed off at speed when the 2nd shot at it hit its skirts. But a minute or 2 later it came forward again, on the same path, only to reverse again when that zook team's last bomb bounced mostly off. At its third advance it did try and skirt away from the couple of hedgerows where my infantry were waiting (rather than sticking to the road as it had before, which ran next to those Bocage), but couldn't keep enough distance and advanced anyway to where the second zook team nailed it. So it did sort of seem to remember that there was a threat there, but its overriding imperative was to get forward (even though if it had made it through that little gauntlet, there would have been nothing there for it to fight, as I'd already passed the point it was probably going to try and defend, not to mention flanked and eliminated the infantry it was advancing to support). So if you get "inside the decision loop" of the plan (though calling it a decision loop is a bit generous - perhaps "if you proceed other than according to the AI plan creator's expectations" would be more accurate), the TacAI might be forced to be less competent than it generally is. Basically, it's still really tricky to get the AI to do anything that involves maneuver without it becoming vulnerable to the unpredictability of the human player.

[John Malcolm]

In a quick battle tonight (infantry only) I watched the AI push it's entire force down one side of the map, through hedgerows under HMG fire, followed by HMG fire + Mortars, followed by HMG + Mortars + Rockets.... they just kept coming.... there was no Plan B.

[hoolaman]

In a quick battle tonight (infantry only) I watched the AI push it's entire force down one side of the map, through hedgerows under HMG fire, followed by HMG fire + Mortars, followed by HMG + Mortars + Rockets.... they just kept coming.... there was no Plan B.

Yep this is my usual experience even with the AI even in scenarios. Infantry especially just RUN. Doesn't matter how many piles of corpses there are for them to trip over, if you have caught them on the move it's like cutting down lemmings.

As for the TacAI remembering contacts. They will re-pick up a contact faster if there is a ? nearby, and they will also directly remember a target for a short period.

[Statisoris]

I once set up a quick battle with a + adjusted US infantry only attacker vs me as German defender. The map was one where there were two bridges for the US to cross. They used human wave attacks the entire game. In the end I think 3 HMG units, 2 crack snipers and a squad of infantry had over 200 kills between them after being continuously resupplied with ammo.

[LongLeftFlank]

My biggest kvetch with the AI other than the absence of event triggers is that units in the same AI group (and there are only 8 of these available) randomly pick their next objective square from those painted in the Order, even if it means a high risk lateral move across the enemy front. This makes advancing in a skirmish line on a broad front, for example, basically impossible unless you have the group converge on a narrow objective area.

In CMSF, my workaround was to use a "jellyfish" model of movement where the AI group would move to a dispersed set of objectives then to a tightly clustered set (preferably out of likely enemy view), then out again, etc. No idea yet if that'll work in CMBN.

FWIW

[Paper Tiger]

I suspect a part of this Lemming behaviour could be eliminated if units did not behave so bravely when they come under fire. If a squad with NORMAL morale takes a casualty, perhaps, it would be better if its current movement order was cancelled in the same way that a HUNT order is cancelled when the recipient spots an enemy unit. Units with higher morale could continue for a while longer with Fanatics rushing head-long into the maelstrom. This would make the AI opponent a little less likely to keep running forward blindly, or at least add some time before the game issues them with their next movement order to get to their group's objective.

I know that when I'm playing, I will cancel a movement order when I see units literally running into trouble. (On those rare, RARE occasions when I played WEGO, I would often shout at the screen 'Stop running, you idiots!' when this happened.) Of course, I will then issue my units with a new and entirely different movement order appropriate to the new tactical situation. The AI won't be able to do this.

[sburke]

I know I've made some very poor decisions recently, but I can give you my complete assurance that my work will be back to normal. I've still got the greatest enthusiasm and confidence in the mission. And I want to help you.

[mjkerner]

Thanks, HAL. What could go wrong.

[womble]

I suspect a part of this Lemming behaviour could be eliminated if units did not behave so bravely when they come under fire. If a squad with NORMAL morale takes a casualty, perhaps, it would be better if its current movement order was cancelled in the same way that a HUNT order is cancelled when the recipient spots an enemy unit.

Sometimes this would be good. Other times it could lead to awfulness, if a casualty was incurred just before the element got back into cover, for example, we'd be as disappointed that they stopped and hit the dirt in the middle of a kill zone rather than extricating themselves expeditiously. It's a thorny problem. Splendid though it is to have a human input to the AI's actions, plans definitely have room for future improvement and innovation.

[Chris Ferrous]

I'm still working on my first, reasonably small, fully fledged scenario (Beyond Briquebec), having done lots of 'little tests'. *still need some playtests so if anyone is interested please contact me*

My experience with the AI hasn't mimicked what has been described here, but I can see how very elementary 'programming' of the AI could lead to the massacres described.

HvH of course presented no problem; make a map, paint reasonable set-up zones, pick a reasonably balanced OOB for the tactical situation, and away you go.

Hv AI defender wasn't much of a problem either and the 8 slots and 6 plans were enough to do the job, and provide a bit of variability. The only problem was whether or not to create different time limits (and timelines for the AI) for We-Go and RT play. I did notice though that any movement by the AI is either done at some speed or at a crawl.

H as defender v AI attacker has had me stumped for nearly 2 months now. I nearly got it right first time, but contrary to the massacres described, the AI WOULD NOT advance any further after becoming brittle, so they get halfway or part way into the objective and then just sit there till their ammo has gone. I've tried second and third waves but when they get to the sticking point they stop too!?

In this regard the 6 plans is ok, but only 8 slots is really limiting options.

To prevent AI 'random spread' I keep the painted target move zones fairly small and directly in line of the line of march I want them to take. Thus any criss-crossing of teams is only within that platoon's operational zone. This seems to work. If the plan needs a flank switch, I've taken them back to cover, and moved them across the rear effectively to advance again on the other side. This seems to work.

Overall though it is a real problem to get the AI to move aggressively effectively.

Again, overall, I'd judge the attacking AI to be hardly any better than the objective (flag) controlled auto-generated AI plans of the old CM1. And for the designer CM2 plans are a lot of work!

[hoolaman]

Again, overall, I'd judge the attacking AI to be hardly any better than the objective (flag) controlled auto-generated AI plans of the old CM1. And for the designer CM2 plans are a lot of work!

Yes I agree. And the amount of work involved makes it significantly inferior imo. OTOH of course the AI plans can do much more interesting things when they work, like aggressive flanking.

I have never observed an AI unit stalled by "just a few" casualties. They more often get themselves wiped out when on the move. Some sub-routines were added for the static defnders in CMSF, like the "bugging out" manouvre, but we need much more of this sort of thing before the AI is even a little bit plausible.

My dream is a little bit of CMx1 in the CMx2 AI and a lot of tweaking of how units behave with the various plans and dispositions. Even the AI tools we have right now could be working so much better.

[Cymru]

I suspect a part of this Lemming behaviour could be eliminated if units did not behave so bravely when they come under fire..

Sometimes the behavior is medal-worthy. In a game of Cats and Dogs one US jeep got hit early on. The driver started to run down the road, and continued over several turns, until he reached the far side of the map and presumably scored victory points for his side. During that time at least eight German units fired at him in the early turns, then no one fired after he got halfway across the map (maybe a silent tribute to his determination).

[altipueri]

Didn't something like that happen at Anzio with a jeep patrol getting all the way to Rome?

[Chris Ferrous]

I think they may have been journalists?

Certainly journalists linked up with the Salerno beachhead before 8th Army got there, but then again one vehicle with non-combatants aboard is a bit different from moving a whole Army!

[RockinHarry]

Once I started working with AI plans 2 month ago, I wondered that these are not supplementary to any strategic AI, but instead are exclusive. There is no strategic AI, like in CMX1, thus the complete inflexibility and disregard for a scenarios goals.

I´d intended to use AI plans (for an attacker) to lead to assembly zones and then let the AI do the attack and capture objectives, but apparently this does not work in CMBN. The AI player does not really know about what to "capture" and with a given "terrain objective", it would never move there on its own, unless just told to move to the area.

So all the "attack", "advance", "max assault" ect. modes is just a "movement to" behavior, but not a "tactical" guideline to capture given objectives. And "movement objectives" is not equal "terrain objectives", so with the lack of the AI´s awareness for a scenarios "goal", I have doubts that at least an AI attacker can be given a halfway realistic attack plan, just by means of applying group movement orders.

Beside end game scoring purposes, do have (Mission) Parameters, Terrain Objectives and Unit Objectives any influence on AI behavior during game?

[George MC]

I think one of the pitfalls with creating AI Plans is desgners over engineering their plans.

I think the CMX2 designer created AI attack/defend plans is, if used well, an improvement on the old CMX1 AI. In CMX1 you took the flags and the AI attacked - pretty predictable, even if it meant the AI leaving good positions.

Now in CMX2 you can have the AI hold the key positions and with a bit of testing have it atatck areas you reckon a human player may well take. The trick is not to too tightly mark out the AI attack plan.

Second tip is if your AI units morale levels are set low then after taking casualties they may well stall (just as units taking heavy casualties would do in RL). How do you get around this? Well you can up the morale levels but care needs to be taken with this. The other option is to make the AI Plan reasonablly flexible so if the AI encounters trouble it can re-route. Not awlays ideal but that's what testing is about.

I'd like to see triggers for sure - it's on the list I believe. But in the short term and as an old boy who had been a miner once told me "you can only piss with the cock you have"! So until triggers etc we have to make do with what is in the game.

Just don't try to over engineer, give the AI some room to adjust and test. The more scripted you try to make your plan the more frustrated you as the designer are going to become.

[Chris Ferrous]

George MC: Are you sure that the AI will try a different route on its own initiative? My comment below summarises the extent of variability in my experience.

Rockin Harry: I have seen the AI make spontaneous repeated minor adjustments to position if threatened, but I'm talking about maybe a 40m radius of action from that 'suggested' in the AI plan.

[womble]

George MC: Are you sure that the AI will try a different route on its own initiative? My comment below summarises the extent of variability in my experience.

Rockin Harry: I have seen the AI make spontaneous repeated minor adjustments to position if threatened, but I'm talking about maybe a 40m radius of action from that 'suggested' in the AI plan.

I've definitely seen the AI deviate by more than that from whatever its inital approach was. In the StuG-fighting anecdote I posted early in the thread, the last StuG was a good 80m off the road it had initially tried to come down. It would have been further, I think, but there was a hedge in the way It was far enough off that I was a bit concerned as to whether the zook team would be able to kill it with their remaining 3 rockets.

[Chris Ferrous]

Womble: True. That's a good illustration. Not seen that myself though.

I'll work on the 'over-engineered AI' concept and knock a few halt, re-group, and shoot zones from my AI plans to see what difference that makes.

I suspect it'll produce the human wave massacres previously described but I'll keep an open mind.

[George MC]

George MC: Are you sure that the AI will try a different route on its own initiative? My comment below summarises the extent of variability in my experience.

From what I've seen in my own AI Plan testing is it will try a differant route. How differant that route is does depend on where the next AI Plan waypoint is and the terrain the AI unit is travelling through. if you want to se this in action then use a very large map and give the AI Group one AI Plan waypoint at the toher end of the map. If you play with timings and 'behaviours' you wills ee the AI do differant stuff. Add in enemy units and the AI becomes a bit more unpredictable - but I think if the desgner can harness that unpredicability to create that unknown element you are onto a winner. It's the Holy Grail of AI Plans

I guess the point I'm making is designers should consider making their AI Plans less scripted as and when they can (there are times for sure where you want to have lot's of waypoints pretty close to each other when you want the AI to do a particular task). Having the AI operate as a human player would is asking too much, so closely scripted plans trying to mimc a human's play style is a big ask I think).

Bottom line is AI Plans are more efefctive if designers work within the limitations of the game engine rather than fight it and AI Plans are more art than science I reckon

[George MC]

Just before we go all dewy eyed over the AI in CMX1 - human wave massacres were still common with the AI...

@Chris - I'm not sure it'll stop the human waves but I guess what you are after is making them bit more unpredictable. As I said I think it's more art than science

At the risk of teaching ya to suck eggs, some variables to work with are:

give the AI more time to get from Point to B to Point C etc

play around with the 'behaviour' e.g. quick will see the AI charge on, advance it'll charge on but with a bit more caution. Maximum assault will see it stop and engage with all guns blasting.

Hope this helps.

[LongLeftFlank]

I know I've made some very poor decisions recently, but I can give you my complete assurance that my work will be back to normal. I've still got the greatest enthusiasm and confidence in the mission. And I want to help you.

LOL! +9000!

[Georgie]

How close to real life is the AIs stubborn advance into fire. Some times probably pretty close. The real life platoon advancing under fire didn't have the knowledge that we the player have. The number of casualties that they were taking would in a lot of cases not be known until the battle was over. Fog of War you know. They were trying to carry out an attack order that they had been given and to stop the advance and report back, hey they are shooting at us up there, would not have flown very well. At least that's how I view most of the AIs behavior. My impression of what a soldiers experience in battle is that it is very focused on just the immediate threat that he is part of and even a platoon leader would have a hard time knowing exactly what is going on until it was over and then it would't be exactly like he thought it was. If the AI enemy gets chewed up pretty bad then it was because of a poor attack plan which happened many times IRL.

[Chris Ferrous]

George

Thanks for the tips. I've tried those though.

I did do a test early on which demonstrated (with vehicles anyway) that each AI behaviour DID have a dramatic effect on how it progressed, but not on the route it took, but maybe the bocage map didn't really offer any realistic alternative.

One thing I did notice, with infantry, was that max assault caused them to move forward in 'leaps' just as in the assault order BUT by crawling only! This certainly reduced casualties, but once the whole group were off the start line for that movement it actually reduced fire output AND tired them out as well. This meant even if they were successful they were too tired to doing anything afterwards.

Whereas Assault simply mimics the assault order.

Advance seemed very much to be a 'carry on regardless' type of order resulting in casualties until morale causes a reaction.