From Basecamp - An early discussion about AI

[Pat]

NEXT!!!

Here's a long running discussion on the AI. This is from early on in the AI's develoment and it has changed and improved considerably since then but I wanted to illustrate some of the back and forth with a few of the testers. Couldn't of done it without them. Again all the AI issues discussed here have been addressed in the full release.

Note: I've naturally had to edit out some top-secret stuff.

von Rundstedt delivers

Paul Dobbins - Playing Build 171, I got in a big Kesselschlacht with vR. I took a chance and encircled 4 PzDivs in the center of the map. The Germans got a big push, and blew past the front line. But the Allies acted aggressively to cut the buggers off. I had them isolated for two days, but just couldn't hold out. Counterattacking Krauts pushed the Screaming Eagles out of the circle, and the Pzs came back to life. Not pretty thereafter, as everything had been committed to hold the Kessel.

Jim Venn – Played a couple of RttM games under build 171 vs von Rundstedt and he needs more work.

The first game it became apparent he wasn't going to make the Meuse, but on the 18th he never ordered an attack on Bastogne despite having 2 armour (one of them elite) right on the doorstep. He instead shuffled troops along the Eastern sectors that were of little consequence.

In the second game I decided to abandon Bastogne completely as the Allies to see what would happen. I halted von Rundstedt in the North so a draw was again the best he could hope for. On the 18th with Bastogne open he moved an armoured unit into and through Bastogne. With units in range and yet to be activated I could have retaken Bastogne which would have left his armour out of supply and the Axis with very little to attack Bastogne once more.

In the RttM scenario I think more importance should be placed on taking and holding Bastogne, especially when moving out serves no real purpose of the Meuse is out of reach anyway.

Jeff Dougherty – von Rundstedt hasn't been tweaked yet for RttM. Right now all the work I've done on him has been for the long game. I'll update you when the next build gets posted.

Jim Venn – Ah ok that explains it.

David Hoeft – Another BoB vs. von Rundstedt- AI seemed stronger and played well; less passing and fewer dubious moves. Game ended on the 24th on points.

It still seems to become less aggressive after the 22nd, with the exception of reacting to getting any AI units surrounded- then it attacks vigorously to reestablish supply.

Paul Dobbins – Von Rundstedt seems pretty formable lately. I'm not sure whether I'm seeing a run of good Axis dice rolling, or the to hit odds have been tweaked, but my units have simply vanished in the face of the 1SSPz and the 116pz. I haven't seen attacks this consistently devastating by single units prior to this, so I'm not sure what to make of it. I'll try again and see if I can both preserve units and keep the river line safe...

Jeff Dougherty – Trust me, we havent touched the combat model. Von R did get a smarts upgrade in tonights build though.

Paul Dobbins – Played again, my side appeared to be doing all right, (i) nothing adjacent to Meuse, (ii) solid defense line, and (iii) holding Bastogne, as per the scenario Briefing: Objectives. But I lost really fast on points. Whew, looks like the vR "fix" is really in: increased attack effectiveness vis a vis the defense (I don't know if it's symmetrical, or just bolsters the AI), and much lower VP threshholds. So it's a new ballgame.

David Hoeft – BoB vs. von Rundstedt (build 172)

Overall continues to improve, but a couple things of note...

Allies has 7th Arm (5 pips) and CCB/9th Arm (2 pips) in Trois Ponts, and the AI chose to use the commandos against the CCB/9th.

Also, the AI left Panzer Lehr and FuhB on the east map edge, unmoved, until late on the 19th, while many lesser units were shuffled forward.

Game ended on the 21st with 15 VP's.

Ros Hermans – 3 BotB games vs vR and I'm pretty sure he's not quite ready for prime time. Two end on the 20th with 6 and 8 vp. It's better, but not there yet.

Often passed when multiple Pz divs were on the east edge of the map.
Attack backward when small units like the cav in the north sneak behind. 1st SS should move backward for that.
Still cut off too easily. All three games saw the loss of at least two Pz div to isolation.
Still allows fairly strong lines to for in the first few day. Just doesn't seem the have that early punch.

Paul Dobbins – I may been playing RttM instead of BotB in the last reported game. I have a Q. Are guys playing with fast AI moves on or off? Anyway, I have had regular success cutting vR off, but counterattacks, especially first thing in the morning before Allied replacements have been tough. Just got nipped again on points. I had recaptured Bastogne and I had a bunch of Germans in the bag again, and they weren't getting out this time; the Germans not bagged were pulling back.. Is there any interaction between supply status and VPs for objectives other than the Meuse? Interesting that vR fights an historical battle towards Huy and the west, instead of north to Herve. I have had some difficulty stopping Pz from getting through the front because I have been preserving my initial armor as much as possible turn 1, I'll rethink that.

Tom Kassel – R (build 173) loses on the 22nd with 7 vp. Many poor German moves. Three pz divisions never moved on the 16th. On the 17th, one pz div chases the cav who had moved to Stadtkyll and then wastes time attacking 106th div. 1 SS reaches Huy on its own and is easily cut off (not quite killed - Germans surrendered first). Reinforcing II SS corps wastes time attacking bypassed 2 ID. Unlikely attack on Verviers (defended by two infantry) by two German infantry has predictable result - heavy German losses and units destroyed on allied turn. Later the same attempt with a pzg and inf - same result. Bastogne (with 10 AD and 4 ID) attacked but not dented a couple of times, then surrounded. Once surrounded, all the German units left, leaving open areas through which Bastogen garrison could put itself back into supply. St Vith was the only VP location captured by the Germans.

In earlier builds, there was an option to continue beyond a VP decision. I think many players would enjoy pummeling a defeated opponent for a few turns, as pointless as the exercise may be.

Tom Kassel – next game v R defeated on the 21st with 18 VP. Again II SS Corps wastes effort against 2 ID and one division doesn't even bother to exploit, remaining in Monschau for the rest of the game. With three Pz divs in Werbomont it attacked Marche, Huy and Ouffet with feeble results and two of the divisions were immediately put out of supply, though one was destroyed before it noticed. There were some curious backwards moves by units on the south edge.

Nicholas Karp – In earlier builds, there was an option to continue beyond a VP decision. I think many players would enjoy pummeling a defeated opponent for a few turns, as pointless as the exercise may be.


Yep, should add a switch for that. Not the first release though.

Tom Kassel – v R dies on 20th, 11 vp.

Jeff Dougherty – All,

Build 176 was just posted with an improved version of von R based on Ros' and Miguel's input. If you're evaluating the AI please download and let me know what you think.

Jim Venn – Had a couple of games vs vR on the RttM scenario under build 176.

I can only assume vR has been tweaked for the long game not the short as I didn't notice any change in terms of realising that the 18th is the last turn therefore there's nothing to lose in attacking Bastogne whatever the odds.

I did notice that he played more conservatively presumably in reaction to the comments in the ease at which he overextended and allowed his forward spearhead to be isolated in the long game. Only trouble is if he's reined in too much he'll find it difficult to win the short game.

Jeff Dougherty – Unless I indicate otherwise, AIs are being tuned for the long game. Please do not test them on RttM unless they have been adapted for it. They will not do well.

David Hoeft – Vs. build 176 von R...

Allies win at end of the 19th with 11 VP.

Commando strike again on one of the 2 pip CC units when there were two 5 pip Allied armor units adjacent that would have been better option.

Pz Lehr and FuhB again sat on east edge until late on the 19th.

Key loss was when 3 German armor units allowed themselves to be cut off in Verviers, went OOS and were wiped out- this because AI ignored the bypassed Allied units in Monschau and let them close in behind the advancing panzers.

Paul Dobbins – I just downloaded the latest build, 176, and there was my good friend vR again. This time those Axis Pzs seemed to be hitting with feather pillows. I dunno, I went through a couple of games (earlier builds) where the Pz's just ate my units alive, especially the dynamic duo of 1SSPz and 116Pz. So it looks like a run of good random number draws for vR did me in. With this build, I noticed a more cautious approach by vR, a methodical massing of armor, but his attacks just weren't generating many hits. Apart from an attempted long swing around my southern flank, there was no penetration past the MLB That had come together by the 17th, though a massive Pz spearhead did manage to drive the remnants of 3 armor units out of my favorite area, Trois Points. I was surprised Dec 16th ended very early, at first I thought it was a crash, but vR passed fairly early in response to an early American pass to end the day. End of the game on Dec 22, with vR at 22 VP and going down....

Jim Venn – Understood Jeff.

Oh and I forgot to mention the new combat animations look fantastic!

Paul Dobbins – I second Jim's comment on the animations. Everything is looking terrific. The end of game scoring summary is a real kick, watching the VP graph unfold with the scoring events popping up on the right margin. I'm having a lot of fun playing this game

Jim Venn – Played vR on BotB and he lost pretty decisively due to allowing units to be cut off. Admittedly I played pretty aggressively and the dice were leaning more towards my side but the 2 fundamental mistakes were allowing 2 armour divs to be cutoff in Werbomont at the start of the 19th, and then 2 Inf plus 1 Mech in Houffalise 2 days later trying to rescue them.

A lesser out of supply was when he advanced 2 Inf in behind Bastogne in Sibret.

vR just squeaked through the 21st on 19VP and was finished on the 22nd with 0VP.

Jeff Dougherty – Once again, new build with improved von R, feedback requested.

Paul Dobbins – I played vR once more before the new build went up. I've been listening to the words of General Venn, I played very aggressively, attacking east, ignoring the Pzs moving west, but cutting off them off when possible. Result, though not as impressive as Jim's, was fun and effective. I think vR may have run up close to 40 VPs before I cut them down to 21 a turn or so later, ending the game. I aggressively attacked the isolated Pzs and killed many of my past tormenters. I could win using my previous careful maneuver, game of positon, but an effective German attack could throw that strategy off. "Back at you" works better, especially by cutting enemy LOS as far east as possible.

Mike Kent – Just finished BotB as allies vs vR. Won on the 23rd with 18 points. AI was keeping it interesting. Seemed like I always just barely had a unit to reinforce or plug a gap. Lehr was down south and moved backwards to take an empty territory. Lost and regained Bastogne. I was able to take it back with the southern reinforcements of 5th inf 26th inf , 80th inf , and 4th armored. AI left the way clear all the way to town.

I hope my recaps are useful.

Paul Dobbins – My experience as well, games vs vR are fun and always there are close calls.

Mike Kent – Just had a draw as allies vs vR in RttM. I believe that is my first "loss" to the German AI for that scenario.

Paul Dobbins – Played vs vR new build 177. Beat him, he had 12 VP end Dec 20th. vR very methodical, less aggressive, but Pzs maneuvered closely and massed for attacks. Two Pzs elim by attacking duo AB divs, 101 and 82. PZs had won an earlier battle vs 1 ID, weakening them, but still surprised the AB crushed them both in one attack. vR threat collapsed thereafter. Will tackle Miguel next.

Tom Kassel – vR 177 loses on 22nd with 12 vp. Better in some respects but still badly flawed. Attacked 106th div on the 16th!?! When the German attack at Losheim failed to kill the cav, I tried moving it to Malmedy joining the 9A cc - bad idea, when 1 ss attacked they retreated to Eupen rather than Trois Ponts. 1 SS P moved to Spa on the 17th, sat for a day or two, then moved out of supply and was killed. Later 9 & 12 SS P attacked a weak force in Werbomont and BOTH!?! used breakthrough move (to Huy & Marche) where they were trivially put out of supply. Only had time to kill one before German surrender. Failing to check supply implications of breakthrough moves is a killing flaw.

In the south 2 PZ spends many moves marching round the south of Bastogne, but this takes far too long and can only be useful if Hoffalize is securely held (which was not the case). At some point Lehr had moved to Manhay where it was easily put out of supply and destroyed.

Mike Kent – I have also seen the German commando played on either a 1 pip cav or 2 pip combat command. It was especially questionable when the location also contained a full armor division.

Tom Kassel – Lost to vR for the first time - VP 62 on the 24th. I have got used to treating his out of supply units with contempt but this time he managed to restore communications, while cutting off some allied units. He managed to take Bastogne, Spa and Verviers while killinga few units during one turn of happy time.

A Meuse crossing was never threatened and at game end the Germans were only just clinging on to Verviers and Bastogne. In the following turn there would have been enough German casualties to swing the VP balance, but it didn't quite happen. A German win but the end game message is totally wrong as Antwerp is under no threat. Something about allied plans being disrupted would be more appropriate

Ros Hermans – Interesting. Ive played 2 games of vR on 177. His opening moves seem to have improved, but I still saw a lot of panzer paralysis. They still get stuck either on the east edge or in a staring match across the Ourthe with alarming frequency. I also rarely see the AI threaten or take Spa. I makes it easy for Allies to ring the Ourthe for the great traffic jam at Werbemont.

Commando behavior is still about sqiurrelly.

Tom Kassel – Next vR is lost by Germans on the 21st with 15 VP. They passed as their first move of the day and so did I. They were correct to pass - there was nothing to do. In this game, I had a solid line Eupen-Spa-Trois Ponts-Houffalize-Bastogne-Longvilly-Wiltz-Beaufort. 2 Pz and two VG divs west of line out of supply.

Jeff Dougherty – Yet another stab taken at von R. At this point, I think I've done all I can with him for our first release. The supply heuristics need to be completely reworked, but that's not feasible within the time we have. It'll have to go into an early post-release update.

(Note: The supply heuristics were indeed reworked and are actually in the first release)

Ros Hermans – Feels better Jeff. At least at the open. He's still been cut off, but seems to be making better choices. Less pazner paralysis. HE still will chase small units the wrong way with big stacks of tanks, but from one game it is improved.

Jeff Dougherty – I may do one more change. Turns out that when I redid his December 16th instructions to make sure he got the most out of the day, I left the randomness low for all subsequent days. That may explain some of his poor performance, although frankly I don't think we're going to see major improvements until Miguel gets a chance to rewrite the watch-your-supply-dummy code.

Ros Hermans – Ok. I'll keep banging on it if only to be able to commincate it's weaknesses once live. Are you going to make the commando change too?

Jeff Dougherty – Don't know. We'll do it if there's time.

Jeff Dougherty – Miguel just sent me some updates too, so there might be one more iteration of von R.

Ros Hermans – Roger. Just wanted to know if we had to test for it. Will await the next build.

Jim Venn – Played a BotB vs vR and echoing the sentiments that the opening moves were much better. Still able to cut 2 armour units early on which cost him badly and won on the 20th with 9VP.

The main problem was he tried to perform an end run around the Southern flank of Bastogne but not taking Bastogne itself makes reaching the Meuse via a Southern route long and slow and he's always going to be in trouble avoiding an early defeat if he doesn't accumulate enough points elsewhere.

Id like to see the AI placing a little more importance on taking Bastogne rather than trying to flank it which I've seen it do a few times now.

Jeff Dougherty – One more version of von R. You can play RttM against him but I don't recommend it. Going to do more work on that tomorrow.

Tom Kassel – I played RttM against vR (180) as the note in the app said both long and short scenarios. He didn't seem to have any idea of the objective and didn't come anywhere near the Meuse.

Then the full game. German surrender on the 21st with 17 VP. 1 ID hamstrung on the 17th so the usual river line was vulnerable, I moved 10 AD to Houfallize to shore it up so 1 SS tries the river crossing into Trois Ponts. Held for a turn by 9 AD CC, then an inf tries and held again. Next turn 116 Pz joins and they duly crush the allied unit so BOTH pz divs exploit to Manhay, leaving a 1 step inf to hold the supply line. 7 AD immediatly chases him away and effectively game over.

There are a few German attacks into forest towns held by pair of allied divisions - usually it doesn't even need all the casualty absorbtion available to emerge unscathed on the allied side.

On the 17th , a pz div assaulted 106 at Schonberg - this is a pointless waste of scarce resources. AI seems to like killing units, even when there is no hurry and there are much more important tasks. A very curious move on the 21st - with 2 ss, 9 ss and and inf adjacent to Eupen, one pz div moved to Eupen attacking 8 steps futilely, the inf and other pz div remained in place in Elsenborn.

Tom Kassel – vR, build 180,18th. Two Pz divs out of supply as usual,but allies stretched. Eupen, spa and Trois Ponts strongly held,a single allied div at Manhay. Houffalize and Bastogne empty. 2 & 9 ss Pz at Malmedy. Moving both to houffalize would be strong with attack on Manhay following turn. Instead, a couple of inf divs move to Houffalize and late in the turn Malmedy is activated. One goes to Houffalize where the force will be too weak to attack usefully but even more ridiculous the other goes to stadtkyll. This must be some flawed internal mapping of the AI terrain model as there is no remote way such a move could be useful. I have seen the odd inf div move to Stadtkyll in the past which made no sense. A mapping bug seems plausible.

Nicholas Karp – Thanks for the detail on the AI, guys!

David Hoeft – A couple of plays against 182 von R...

First game the AI picked a 1-pip 99th Inf for commando target (though in the second game it took the 7th Arm).

Attack woes- a suicide attack from Losheim to Malmedy- the AI attacking with a couple of 2-pip infantry against a total of 8 pip elite infantry. Also, understrength attacks that could have been done better with no risk- as when one unit left Bastogne to attack a lone Allied unit in Neufchateau, when all three could have safely left Bastogne and joined the attack.

AI did make one very nice shifty attack, leaving (with 3 armor units) a contested area where it would have been a hard fight and moving instead to a weakly held non-adjacent area.

Less trouble now with AI leaving units on the east map edge- things seem to move forward more reliably now.

Finally, armor doesn't seem to stick itself into a noose as much, but it seems doesn't concentrate well or advance as aggressively as before.

It'd be very interesting to see how a brand-new player would do against von R right now, though...

Ros Hermans – I'm a few plays into 182 now. Progress has been made on the first two turns. It still will over extend and follow with moves to cover supply lines and seems to really love north east corner. Armor still can get caught there and AI will move inf unit up there rather than move west.

Overall I think it's pretty much in the same boat as Monty, maybe a bit worse. It likely will run ok against new folks but my sense tells me experienced wargamers will figure it out fairly quick.

Jeff Dougherty – David's going to send out another build soon. No more changes to von R, but I've got an Axis agent that plays RttM acceptably now. It's heavily scripted and probably quite brittle, still does some stupid things in the south, but it'll head for the Meuse and Bastogne without too much exposure to being cut off. It'll work for the short term, Where the short term is defined as "the next week-ish". I didn't want to go the scripting route, but with the AI's current lack of supply awareness there just wasn't another way.

(Note: the scripting was temporary to circumvent an AI feature that at the time was not implemented. It was initially done this way so the testers had a base to work from)

In the medium term, which means either for the first release or for an update coming out shortly thereafter, Miguel is going to work on improving the AI's supply-line system so it can get a better idea of when it's in danger of being cut off. That should solve a lot of the problems with the von R AI for the long game, and let me write a more general RttM Axis AI.

Onwards.

Ros Hermans – Cool.

[jeffd]

Jeff Dougherty – Yet another stab taken at von R. At this point, I think I've done all I can with him for our first release. The supply heuristics need to be completely reworked, but that's not feasible within the time we have. It'll have to go into an early post-release update.

For the record, I'm very glad that I was wrong here. At the time we were thinking that our first release candidate, which had substantially less polish, might have to be our initial version. Fortunately, Apple approved the second version much more quickly than expected and we were able to release with that.