AI vs AI game

[henri51]

I decided to let the AI play both sides overnight (it takes about a half hour per move) to see how the AI would make out.

When I got up this morning the game had not crashed (yay!), and the Soviets had achieved an overwhelming victory. [X(]

The losses on both sides were comparable, but the soviets had about 9000 victory points to the German about 2000. Except for the border main objectives (Orsha, Vitebsk and Mogilev), the Germans had captured only Demidov, and Metislavl, Smolensk, Yartsevo and Roslavl were all still under Soviet control, although Smolensk and Mstislavl were all but surrounded and probably ready to fall.

This result indicates that either the game is unbalanced or that the AI is much better at defensive play than offensive.In the only AAR published so far (Keating vs Alston), the result was a draw, but I noticed in that game (in the Armchair General I think, but available on the SSG site), that there were no encirclements to speak of, whereas in the real campaign, there were a bunch including the huge one in front of Vyazma where the Soviets lost half a million men.

Since I did not watch the play, it is difficult to say what "mistakes" if any led the Germans to this dismal result, or to what extent it might be due to any design or programming deficiencies. But I have to give credit to SSG for allowing the AI vs AI play,because most games do not allow that - maybe to avoid making AI deficiencies too obvious.

Opinions welcome.

Henri

[JSS]

Just one comment on the AAR... Supreme Commander Tempest (Soviet commander in AAR) is perhaps the toughest opponent going (without question for the Ardennes Offensive and ATD games).

If you can match his performance as the Soviet commander, then consider yourself elite!

JSS

[JudgeDredd]

I don't intend to play AI vs AI - so am not concerned about the results. Clearly when doing that one is affected by the others decisions and movements.

As for the AI being better in defense than attack, I thought it was already understood that an AI performed better in defense than attack.

So your findings don't really concern me too much. I'm sure the AI will do just fine against my severely lacking game style

[Adam Parker]

ORIGINAL: henri51

When I got up this morning... the Soviets had achieved an overwhelming victory. [X(]

Looks like the Germans won't get to Moscow after all! [;)]

Hey I've tried many times to crack through the Soviet lines in HPS's Smolensk 41 (my favourite game of the Panzer Campaigns series). Even vs HPS's AI (not the best) breaking through the Vyazma-Orsha-Mogilev line was not an easy task. Remember AI's just learned to play chess.

[e_barkmann]

you've only run this test once I assume. Run it 50 times and let us know what results you get.

I have been on the receiving end of a German assault several times and I can tell you it is most definitely not all roses for the Soviets.

The salient point here is that I do not expect to see really clear statements of how great this game is for at least another month or two. It will take at least that long for people to actually play the game and work out some basic strategies and observations.

Perhaps it's time for me to go back to Tetris -

This game is deep beyond my ability to describe how so.

cheers Chris

[henri51]

ORIGINAL: Chris Merchant

you've only run this test once I assume. Run it 50 times and let us know what results you get.

It COULD be different, but given the lopsided result, unless the scenario is "chaotic" (small differences in initial conditions make huge and unpredictable differences in the results), I find it hard to imagine that the Germans could win in AI vs AI.

Since it takes about 10 hours to run a game, it could take a long time to find out - but 50 games are not really required (the small sampling number for significant results is around 20 - which is still a lot).[X(]

However it is true that the proof in the pudding lies in human play, so I also hope that we will begin to see results trickling in. Personally I hope that the Germans are able to carry out encirclements and that the game leaves some scope for maneuver warfare. There are contradictory indications for this: in his AAR. Keating says that he did achieve encirclements in previous games (presumably against the AI in alpha and beta versions), but he also says that the best German strategy is to stay adjacent to the Soviet units so that they cannot get replacements, which is not conducive to maneuvering. So time will tell...[:'(]

I am thinking of trying an AI vs AI game with the Germans having a +1 advantage on the dice which should aloow them to obtain at least a draw if the scenario is not unbalanced.[:D]

BTW your video tutorials are very useful, although through no fault of yours, I remained with some confusion about the differences between supply radius, command radius, and areas of operations as indicated in another thread. [:o]

Henri

[JSS]

ORIGINAL: henri51

It COULD be different, but given the lopsided result, unless the scenario is "chaotic" (small differences in initial conditions make huge and unpredictable differences in the results), I find it hard to imagine that the Germans could win in AI vs AI.

Ummm, think AI vs AI is not a good perspective for this scenario.

Its a large scenario and is a truly a matter of strategy... try playing yourself and see how many times a surprise break thru occurs!!! [;)]

JSS

[henri51]

I tried it again (AI vs AI), but this time I gave a +1 dice advantage to the Germans. Same result, Soviet overwhelming victory with an advantage of almost 8000 victory points.

I don't know about surprise: although the Soviet units were mostly half-strength reserve units, they were all over the map and I don't think that it is too difficult for the Soviets to know where the Germans are and where they are headed.

Henri

[henri51]

I did it again since I was going out of town.

This time, I gave the Axis MAX computer advantage and left the Soviets to normal, and left everything else normal.

When I came back the Germans had wiped the Soviets off the map and conquered not only Smolensk but aalso Kirov and Vyazma and everything else, with over 6000 victory points more than the Soviets. [X(]

So there must be an intermediate setting where the game is even.

I agree that AI vs AI may not be a perfect representation of what will happen in player games against the AI. OTOH the AI does not "know" if it is playing against a human or the AI, so it should give one SOME indication of what to expect playing against the AI.

Henri

[TargAK]

AI vs AI works pretty good for chess engines, matter of fact that's how the different engines stack up against each other at the end of the day.

[JSS]

ORIGINAL: henri51

...So there must be an intermediate setting where the game is even.

I agree that AI vs AI may not be a perfect representation of what will happen in player games against the AI. OTOH the AI does not "know" if it is playing against a human or the AI, so it should give one SOME indication of what to expect playing against the AI.

The scenario balancing is based on capability to win. A human player can win (or lose) with either side... the capability to win is there, victory comes done to strategy (and especially your opponent's counter-strategy!) and the two player's relative skill levels.

There's three levels to the ATD2 AI... basic engine as coded by Roger Keating, scenario details as created by Ian Trout, and the AI strategy created by Ian Trout.

Because Ian creates his AI formations and strategy based on how he would play (and he's a very tough opponent!), the AI plays the game as Ian would. So its essentially Ian vs. Ian when you run the AI vs AI test (his plans, his spherepunkt, his defensive vision). What you're seeing is something where Ian's Soviet strategy is getting the best of his German strategy.

So far, I routinely kick the Soviet AI in the arse... and, um, well, I don't do so well against the German AI (games are very close... more to the point, I can't keep the kind of relevant defense in place that I'd prefer to have!). This is because my strategy puts the Soviet AI off balance... and the German AI keeps my Soviet defense off balance.

JSS

[rickier65]

ORIGINAL: henri51

......I agree that AI vs AI may not be a perfect representation of what will happen in player games against the AI. OTOH the AI does not "know" if it is playing against a human or the AI, so it should give one SOME indication of what to expect playing against the AI.

Henri

I agree that an AI vs AI shootout has some benefit, so thanks for running these.

Rick

[Henri]

I ran another test, this time with the German AI "only" at ++ computer advantage. Once again the Gertmans wiped the floor with the Soviets capturing all of the objectives, so I guess that advantage is still too big.

In sum, so far a +1 advantage in dice roll is not enough for the Germans, and a ++ advantage in computer is too much. So I guess that the next test should be a + advantage in computer. Maybe tonight...[>:]

Henri

OK I did that, and the Soviets still got an overwhelming victory, despite the Germans having killed almost 10 times more units than the Soviets did, and the Germans having about 3 times as many objectives including Smolensk (but not Metislavl and Vyazma).

So it appears that for the Germans, aiming for Smolensk is not enough, and they have to go for Vyazma and reduce their losses - a 3/1 ratio of losses will not do [X(]. And if there is a problem (and it is not certain that there is), it is in the distribution of victory points and not unbalanced forces.

[Anonymous]

[Deleted by Admins]