Patch?

[MarkShot]

Here is a very graphic example of the more aggressive AI in action.

I had this Highway covered with a Regiment of tanks delaying backed up by two arty batteries. In 12 hours, the Germans have covered 15km. I have already lost 40% of my armor and my forward arty battery failed to displaced quickly enough and is in the process of being overrun. It looks pretty certain that I can simply write it off. Since its only chance of escape would be at night fall, but that is still too far away. Aargh!!!

The armor delaying action was to buy time for my two static positions in close terrain along the Highway and my MLR to get prepared along with priming a couple of bridges. But the loss of a battery is really going to hurt. Additionally, I was expecting to have some armor left after delaying action to help hold the center of my MLR. I doubt I will. Not to mention if the Germans hit my MLR en masse as opposed to piece meal as I would have hoped, I don't know if it can hold.

This is really bad! I hope the patch gets released soon so that I won't have to be the only one around here to face such humiliations!

[MarkShot]

Well, it has been a grueling evening both intellectually and ego wise.

This is what it looks like when a beta tester loses. This is the First Clean Break scenario. I have battled the Germans in a running fight over two days for 40km.

Finally, they arrive at Day #3 at Sun up at my main line of resistance. A little after Sun up (06:00), my Exit Objective (with auto end) condition goes active. If I can collapse my MLR and exit enough units, I can claim a victory.

Here you see the sad reality. I have exited almost my full MLR and yet I have only met 85% set of the objective's criteria. Why? Because the battle delaying the Germans to the point that I could pull out the bulk of my force has been too costly. There will be no immediate end with me being the clear victory.

[MarkShot]

The Germans five hours later will trigger their auto-end exit objective as they continue to hotly pursue my retreating forces down the Highway.

This will result in a draw for me. (which I had intended to show you, but reloading it CTD's ... oh, well ... just when I was going to certify it with the seal of approval ... I am off to the bug forum)

I know some of you will be dissappointed to read of possibly even more delays. But rest assured, the improvements made are giving me a serious run for my money (although I get this for free). I have four years of experience at this four years already. I think you folks are going to be in for a lot of fun when we finally wrap this us.

[Arjuna]

Yes it's a bummer about the CTD Mark. But I reckon I can fix this asap. It's just a case of an inappropriate error check. On loading the game checks that every task is within the scenario time frame - ie between scenario start and end. But Mark has saved the game in the AAR, which technically is one minute after the scenario end, hence it asserts. This should be a pretty easy fix. No other reports have been entered. So I'm still optimistic we can put out a new installer for testing tomorrow and sign off by Friday.

[Arjuna]

Fixed. [:)]

[MarkShot]

What? After fixing that you didn't hang out my dirty draw for the whole forum to see?

[MarkShot]

There you go. I threw in Dave's latest executable. And now everyone can see my dirty draw. (I just was too beat up to exit what I needed for a victory.)

I have rarely lost playing COTA. This new AI did me in good.

In retrospect, I should have played to delay/hold the Germans and score what points I could exiting while stopping them from doing the same. That certainly would have been a viable strategy. Instead I went for the glory of a decisive victory with an "auto-end" condition. Yes, I was seeking favor in the headlines, but then what great WWII general didn't?

<rats, lunch is over; no more COTA for me>

PS: Dave, if you are reading this, we need an exit point calculator ... meaning select a set of units and see how fares against the exit criteria. Not really important for exit objectives when just scoring points, but essential when "auto-ends" are involved. I suppose it might have just saved me from this embarassment.

[Kabbers]

ORIGINAL: MarkShot
I have rarely lost playing COTA. This new AI did me in good.

I love it There should be a word reflecting the enjoyment of being busted by an AI shouldn't there? If there is one, we all seem to have it here, and I can't wait to plunge deeper into this particular pathology. Thanks for sharing symptoms and signs of the contagion with us until everything is ready [:)]

I was intrigued to read in the posts above a description of COTA AI as an "expert system" ... the wikipedia entry for example of 'expert system' is here: http://en.wikipedia.org/wiki/Expert_system

It made me smile because I have only met expert systems (or though I had met) until now that were medical based, like Mycin (http://en.wikipedia.org/wiki/Mycin does wikipedia have no topic untouched!?) so it's quite mindbending to bend my mind to the idea of an expert system dealing with the COTA!

I am stuck in a 'basic chess ai' mindset, where you just create legal moves, eval each position, and create some iterative search trees and heuristics to optimize route to (hopefully! hehe) the best eval branch, and then make that move. But I still can't quite get my head around COTA AI as an expert system... and perhaps it's best that way LOL, either way how marvellous a thing

Looking forward to it & thanks again for a great thread,
K

[MarkShot]

Well, normally, if I won most of the time, I would be long done with a particular game and move on, but not COTA:

(1) Despite the likely knowledge I am going to win, the reward is in the solutions and development of effective techniques. I've gone from a guy who replayed RDOA scenarios over and over again but could not get a draw to someone who maintains a gaming guide for the series. In the process, I have learned much about ground combat strategy despite a "gamey" tidbit here or there.

(2) All the scenarios have options to reduce reinforcements for your side and add reinforcements to the enemy, weather, supply, etc... Except for MAX order delays, I always play the defaults. Partly I continue to do this, since I believe most players are likely to do this. Aside from writing Mini-Guides and testing, I also write scenario reviews and try to envision what kind of experience the average customer will have with the scenario. However, if I ever had time to play and replay scenarios over and over (not for build version comparison), then I am sure I would bump the difficulty level to up the challenge.

(3) Aside from a really great interface and mode of play, this engine has one of the best gaming engine AI's out there. Someone recently said to me that Dave didn't really need to do all this AI work and delay BFTB; the AI was good enough. In part I agree as I think those who found the AI lacking in aggression were probably better than average players. However, despite any delays, I think perhaps this was a worthy endeavor business wise. When I talk to people privately about the game, more often than not, the thing that makes them a customer now and in the future is the quality of the AI. In short, among discerning customers of war/strategy games it is realized that this engine has a superb AI. I think years ago when I studied some marketting, one would say that the high quality AI is the engine's and PG's key value proposition.

---

Regarding expert systems. I don't know too much about the code that's in COTA. Although occassional Dave launches into C++ speak. Unfortunately, my security clearance isn't high enough to see the code. But given how much code there is, perhaps that's not a bad thing ... if you think 500 pages of game docs is a lot to read already. Well, one thing I know, if it is an expert system it certainly isn't a neural net and is probably either a forward chaining or reverse chaining rule based implementation. Just speculating ...

Catch you later.

[HansBolter]

I am gonna paint a huge target on my forehead here since I have already acquired pariah status by delaying the patch by stating that while I think the AI does an incredible job at the tactical level it is still lacking at the operational level.


Allowing the AI to decide how to execute an assault whether with a battallion headquarters controlling its constituent companies or with a regimental headquarters controlling its constituent battalions is one thing, but expecting the AI to make good decisions about where along an enemies perimeter to concentrate it's main thrust, which regiments/brigades to assign to which sectors of the front line, ie....how and where to concentrate and focus it's effort to achieve victory when there are many disparate and dispersed objectives such as in the Malta scenario and how to read and react to the concentrations and focus of the enemy seems to still be too tall an order to expect from even the incredible AI of this game.

The AI does things tactically very, very well.

The AI doesn't appear to do things operationally well.

While not having played it, the scenario you guys have been using to showcase the revisions to the AI, appears to me to be unidimensional. The AI has a definite, finite, directed objective to exit the board at a certain point and admirably drives relentlessly toward it. I have to wonder if this is the best choice in scenario for testing the AI. The Malta scenario, which I have played extensively, seems to do a good job of showcasing the AI's limitations as an operational thinker.

Please don't get the wrong impression form this post.

I love the game.

I am enjoying playing it immensely.

I greatly respect the accomplishments of the designers.

This isn't intended as a slam or insult to anyone involved in the development of the game.

It is more of a plea from a life long wargamer to those "in the know" in game development as to whether or not there is even a chance, given the current state of AI development, of creating an AI that CAN be effective at the operational level.

[MarkShot]

I think we arrived at these three scenarios for repetitive testing:

Elasson Rear Guard
First Clash at Veve
First Clean Break

for the following reasons:

(1) They cover a lot of ground for the attacker.
(2) The defender is falling back. So, it is pretty noticeable if the attacker either fails to fill in the vacuum left by the defender or maintain contact with the defender.
(3) The attacker seems to have the advantage of fire power and mobility so there is little excuse not to come on strong.
(4) Aside from bug fixing, the task put before us wash that the AI needs to play more aggressively.

I, myself, started my initial testing with "First Clash at Veve". But I discovered that I was CPU challenged so I decided to downsize to ERG.

Initial testing showed the AI allocating too much force to holding already taken rear area objectives and allocating too little force to the objectives on the edge of the advance. Primarily through changes in the code (and to a lesser extent changes to some test scenario; number/type of objectives and scenario duration) the above tendency was reversed. The AI will now press hard and put its main strength on the edge of the advance and only lightly defend secure rear area objectives. I wish I had more time to test, since in a delaying action you want quite the opposite.

On the whole, while testing I found the AI doing better on the attack in such scenario than the defense. I attribute that to the following:

(1) Performing a fighting retreat with a weak and slow defensive force is a more complex problem than blowing such a defense and then exploiting.

(2) I managed to optimize the priming and blowing of bridges as much as a human can given the game mechanics. This made my defense reasonably effective where otherwise I think the AI would have smashed through me. However, the AI on the defense will prime bridges, but it is not going to prime every structure on the map over three days. Maybe just a few.

Additionally, I watched the AI and appears that it has still retained its ability to attack well and find holes against a broad static defense. This has always been one of its strengths which I tend to describe as amobeoid motion. Of course, sometimes it can spend too much time probing rather than just landing a big punch.

In the small, one of the most noticeable changes in the way it plays is that it will spend a lot less time running into opposition and beating its head against a blocking position with a road march. It will more quickly respond to the blocking position and launch an assault to clear it.

I am personally quite happy with how the AI plays given what I have seen in other games. I doubt that it rivals talented human players. But most customers will probably find it a challenge to their human smarts for an extended period of time. Of course, they eventually may reach a point of seeing its weakness as you have. However, it will probably take more than just a few week or months to reach that point.

---

I think the patch results will be pretty good. It is true that only a small number of scenarios were focused on. But consider that we are only a hand full of people and even a small 2-3 days scenario can easily take an entire night to complete when playing carefully watching for and reporting issues. It was the best we could muster. As Dave has said the AI is a "generic system". So, these improvements should manifest themselves across the board even if we didn't retest various scenarios. In the meantime, if there should be an unwanted side effect, I trust it will get reported by the community if we missed it.

{Sorry to write so much ... just wanted to make clear how we approached this.}

[Arjuna]

ORIGINAL: MarkShot
PS: Dave, if you are reading this, we need an exit point calculator ... meaning select a set of units and see how fares against the exit criteria. Not really important for exit objectives when just scoring points, but essential when "auto-ends" are involved. I suppose it might have just saved me from this embarassment.

Huh! So the UI is to blame for your rubbing, hey? They say blame should always be apportioned and as far away from oneself as possible but this is a low note Mark. [;)]

Seriously, I agree with your suggestion. [:)]

TT3156 - UI - Exit Point Calculator for Units

[Kabbers]

ORIGINAL: HansBolter
whether or not there is even a chance, given the current state of AI development, of creating an AI that CAN be effective at the operational level.

What a super post! and I know you seek an answer from those who might or can rather than me Hans hehe, but I did just want to ask (for questions are always fun upon questions) whether (or not, for my own education and indulgence) your question of "effective" operational level AI is really a reframe of the classic AI "frame problem", conveniently defined here for instance:

http://www.byte.com/art/9611/sec3/art12.htm

I definitely don't mean to ask this as a wisecrack, I've never (I think) played an operational wargame (although I'd like to try and I know you want (and probably will get, lucky us) an answer from a wargame peer, but I am intrigued whether 'context' is key to strong "operational" play as you ask? That seems key to your main question.

If 'context' is key, then back to both Markshot and Arjuna's posts on the AI in COTA not requiring "scripts" ... well for an operational AI, how much 'context' (without recourse to scripting) about the situation/campaign/environment etc. does a really "effective" operational AI need? and also - does "effective" mean 'well' or 'like a human being'? I ask because cognitively we are excellent at considering context - but as you point out about the COTA AI, we are often poorer (poor us!) at tactical level 'closed-set' monotonic analysis.

It's nice in chess when the broad 'context' of a plan or 'strategy' (closest I can come to operational in chess? or please correx me) beats-out a tactical genius (Capablanca v Alekhine (e.g. http://www.chessgames.com/perl/chessgame?gid=1270221) comes to mind), but Capablanca had a remarkable ability to 'understand' the context of a position beyond the layer of tactics. Without "scripting" (the other AI method Arjuna and Markshot mention) I don't think anybody in the history of AI has actually solved the AI "framing problem" yet - so technically sans scripting I can say even in my novice wargame ranking that the answer to your question is no (although what fun when they do!). But so far, the poor robot of the framing problem, in a non-monotonic environment, will still fall over/be blown up, unless someone covers that particular situation with a specific script. Open-world AI isn't there yet, and I believe the mathematical analyses, and epistomological analyses of that problem, are still in deep torture hehe

(this is just my little intermezzo, I await more can-do answers to a wonderful question Hans, and i really hope someone has a great operational AI out there, for all of us)

{embarrassed add-on: I can only find examples of Alekhine beating Capablanca LOL but here's a lovely one if you like chess: http://sport.guardian.co.uk/chess/story ... 33,00.html)
Quote: "Capablanca is sometimes called the Mozart of chess. Which must make Alekhine the Beethoven - slightly bonkers and able to conjure up visions that elude everyone else. "[Alekhine's] conceptions were gigantic, full of outrageous ideas," said Bobby Fischer, explaining why the Russian was a dangerous role model. "It's hard to find mistakes in his games, but in a sense his whole method was a mistake."}

[Arjuna]

Well HansBolter you'll have to wait and see for the patch. The prime focus of AI enhancements has been to what I refer to as the "strategic AI" or what you refer to as the operational aspects. I've played the Malta scenarios briefly, admittedly to track down a bug ( so my focus was elsewhere ) but the new startegic AI certainly gives you a better run for your money.

[Arjuna]

Kabbers,
&nbsp;
Re framing problem. I must admit to not spending much time reading about AI theories. I've always been a do-er rather than an academic. I recall a time back in the early 90's when I was consulting for the Australian Defence Scence and Technology Organisation. They had pooled together a building of people all armed with PHDs to explore the use of AI in military simulations and C3I. They were all working on this theory and that. When asked what approach I was going to take I replied I'm just going to analyse the problem and model the thinking processes I believe happen in reality as best as I can. This met with much laughter and derision. It's with a wry smile that I am here today having delivered what some say is the holy grail of command structure AI while none of those fellows ever produced a program that worked.
&nbsp;
The real issue after identifying what decision making processes are made in reality is to then identify what data you need to make those decisions. In other words what situational awareness does the unit commander need to have to make a reasonable decision. This for us has been and will continue to be an evolving processes. When we started we mapped out what we then thought was every possible requirement. This did not survive contact with reality for very long. Writing AI code has taught me that the "let's do all the design work up front" approach just doesn't work. It's very much an iterative process of trial and error. Sometimes this requires you to make major overhauls to the data sets and the processing code. For instance in developing COTA we realised the need to have multiple movement maps based on movement class ( ie foot and motorised ). This in turn required a massive reworking of the movement and planning code.
&nbsp;
Oh boy just realised the time. Got to go and crush a bug. I'll be back. cya [:)]

[Kabbers]

ORIGINAL: Arjuna
Got to go and crush a bug.

Thank goodness for you practical approach
From one-of-the many-who-like-being-busted-by-working-AI

(I'm still wondering if AI COTA is Alekhine or Capablanca)

[Arjuna]

Kabbers,
&nbsp;
If I knew what those terms meant I could tell you. [;)]

[Arjuna]

HansBolter,
&nbsp;
One other thing, we deliberatley gave priority to the tactical AI. Our reasoning for this was to start with a firm foundation. You need a good tac AI to support a command structure. Without a good simulation of the command structure you cannot hope to produce an effective strategic AI without case specific scripting, which we wanted to avoid.
&nbsp;
While the tac AI still has many areas for improvement is now sufficiently powerful to support a command structure and so now we had the foundation to start enhancing the strategic AI. The future for its continued development looks good.

[goodwoodrw]

Well, is tomorrow the big day? to be or not to be![:D]

[Arjuna]

BASB,
&nbsp;
Well the short answer is unfortunately no. We hit some bugs in the multi-player testing, now resolved. I'm awaiting the all clear from the testers. So far ( less than 24 hours since last beta update ) all is good. Autotesters are giving the all clear. If tomorrow morning this still is the case, we will ask Matrix for another Installer build, test that ( and I don't anticipate any problems there ), sign off and then go for public release as soon as Matrix can arrange it.