Civ4 AI Survivor: Season Three

[Sullla]

The writeup for the Championship Game is finished! If you would prefer to watch the Livestream footage, here's a direct link to the footage. Next up will be a Conclusions page with some analysis and rankings of the leaders after three seasons of competition. Look for that coming later this week. Thanks again for everyone who took part in this year's AI Survivor competition!

[Emperor K]

Thanks for reporting on these games. I was only able to watch one of the live streams.

[Sullla]

One more page to add: here's the Conclusions section for Season Three with lots of data and leader analysis. For all the concern in the opening round that these games were completely random, I think we ended up with a group of dominant leaders in the Championship game this year and the playoff games separated the powerhouses from the pretenders. This page includes download links to my Excel spreadsheet and the starting savegame files from all 13 games for anyone who wants to poke around with them. There's one more test that I want to run in the next week or two, simming out the same starting save file repeatedly to see if the games play out the same way or go off in different directions. Just how random are these games if they play out from the same starting position repeatedly? It should be fun to find out.

I hope you enjoyed this season of AI Survivor. Thanks everyone.

[GreyWolf]

I wonder what is the average peaceweight for the leaders that reached the Championship in all three runs? It seems to be on the low side (although it's important to note that the initial peaceweight is always randomized by any number from -3 to +3). Still, only two championship leaders had an initial peaceweight above 6 - the incomparable Mansa Musa and Elizabeth (from Season 1).

Most unit underbuilders (with UnitBuildProb less than 30) don't fight well, with the exception of Catherine, who manages to build enough units due to her ultra-fast expansion, and Mansa, who scored a number of kills due to his high tech rate. Gandhi and Huayna can play the same role as Mansa when the competition is weaker.

Overly aggressive leaders (MaxWarRand=50) are usually too aggressive for their own good.

I do think that some AI's just ended up unlucky in all three rounds.

[Zalson]

Sullla, thank you for putting this together. I really enjoyed it.

I don't know if it's too much trouble, but I'd be interested in seeing those same start simulation games.

[rho21]

Thanks very much for running another greatly entertaining season. The reports are particularly appreciated as I don't have time right now to tune in to the livestreams.

Really good to see the new statistics. I've always thought the old-style ones were a bit unhelpful (but of course haven't been bothered to do anything about it).

A couple of statistics that might be interesting if anyone has time to put them together:

• those same game by game rankings but excluding the performances for the season in question (which as you mention naturally causes a bias towards later games)
  
• a value-added statistic for traits - in other words a way of attempting to filter out the benefit a poor trait gets from the leaders who have it paired with a strong trait
  

[Axiis]

There's one more test that I want to run in the next week or two, simming out the same starting save file repeatedly to see if the games play out the same way or go off in different directions. Just how random are these games if they play out from the same starting position repeatedly? It should be fun to find out.

I hope you enjoyed this season of AI Survivor. Thanks everyone.

It wouldn't surprise me to see quite a bit of variation based on who declares war when and who comes away with the spoils. Caesar in particular is a loose cannon who gets away with his overaggression thanks to his Praetorians basically functioning as an autowin for classical era wars. That said, I just don't see how Mansa makes it through the game without getting dogpiled - the diplomatic situation is just brutal for him.

I wonder what is the average peaceweight for the leaders that reached the Championship in all three runs? It seems to be on the low side (although it's important to note that the initial peaceweight is always randomized by any number from -3 to +3). Still, only two championship leaders had an initial peaceweight above 6 - the incomparable Mansa Musa and Elizabeth (from Season 1).

Most unit underbuilders (with UnitBuildProb less than 30) don't fight well, with the exception of Catherine, who manages to build enough units due to her ultra-fast expansion, and Mansa, who scored a number of kills due to his high tech rate. Gandhi and Huayna can play the same role as Mansa when the competition is weaker.

Overly aggressive leaders (MaxWarRand=50) are usually too aggressive for their own good.

I do think that some AI's just ended up unlucky in all three rounds.

The average peace weight overall is just over 4, but the average peace weight for later round games tends to be significantly lower, largely due to the poor performance of most of the "good" leaders. There are 9 leaders who have never scored a single point in any of the three seasons - never placed and scored no kills. These perennial losers average a 6.5 peace weight, and that's after including Gilgamesh's 2 peace weight. The top 11 leaders by my calculation (there's a 4 way tie for eighth) have an average peace weight of 3.27, which goes down to 2.3 when you take out Zara's 6 and Mansa's 9. Basically, the low peace weight leaders feast on the oft eliminated high peace weight leaders, which makes the playoffs incredibly difficult for guys like Gandhi and Mansa who need another high peace weight meat shield to buy them time.

Darius, Hammurabi, Louis XIV, Peter, Alexander, Augustus, Churchill, Genghis Khan, Montezuma, Ramesses, Roosevelt, Saladin, Bismark, Frederick, Gilgamesh, Hatsepshut, Isabella, Sitting Bull, Victoria, Washington, and Willem Van Oranje have yet to score more than one point in a season.

The top 11 civs don't have a _ton_ in common, but they tend to
1) have a low-neutral peace weight (MM being the glaring exception)
2) not be too focused on units or wonders, defined by having no more than 2/10 differential in their unit and wonder building preference - in fact the only top AIs to have more than a 2 differential are Huayna Capec and Pacal, who both have 8/10 for wonders and 4/10 for units. Incidentally, both of them are financial. However, every top civ has at least 4/10 rating for units.
3) declare war at pleased, allowing for wars of opportunity - the exceptions are Justinian and Cyrus, and both of them have at least 7/10 aggression.
4) be military or production focused - only Mansa, Pacal, and Zara are lacking both military and production flavors, and none of the top civs have a single focus flavor.

Basically, the most consistently successful civs don't go too hard towards either units or culture, but prefer to build strong militaries and are able and willing to eat a weak neighbor when an opportunity presents itself. Mansa is the exception to a lot of these rules, but he's so damn good at teching that he can get away with it.

Of the loser civs, I think the ones that most fit the profile of a potential winner are Louis, who tracks as a combination of Kublai Khan and Huayna Capec; Peter, who only seems to be held back by his lousy philosophical trait; Bismarck, who has the profile outside of having a single flavor (military, so that's not bad), slightly suboptimal traits (expansive and industrious) and being tied to the underwhelming German civ; Gilgamesh, who looks good aside from having the abysmal protective trait; Isabella, whose big problem is her single flavor (RELIGIOUS, duh); and the curious cases of Willem and Victoria. Vicky seems to be held back by her peace weight, because she looks like she should be a top leader otherwise. Willem is even stranger because he checks a LOT of boxes - neutral peace weight, solid traits, good flavors for a potential runaway, has a 4/4 balance for wonders and units  - he seems like he should be Mansa Musa with a more favorable peace weight. Instead, he's never won a damn thing. It's very curious.

Of these, Peter and Louis have the best track record, with Peter in particular always seeming to be this close to doing really well. Bismarck has a habit of getting off to a decent start before being tag teamed by low peace weight bullies, while Gilgamesh has just not caught a break in three seasons. He was trapped on a peninsula in season one, and was in abysmally bad diplomatic situations in 2 and 3, somehow ending up in games full of high peace weight leaders in the latter two seasons. Isabella typically looks strong for a minute, but she seems to have a hard time getting that early kill AIs really need to start a snowball, and her late game never seems to go particularly well. Victoria has a particularly checkered past - she's been rolled by a runaway civ, partitioned by aggressive neighbors, spent the game as a useless rump state, and thrown away a nearly certain win in favor of a truly terrible attempt at a cultural victory. I'm still not sure what's up with her, but she sure seems like she hates winning. Willem likewise remains mysterious, as he's suffered from bad diplomatic situations but also once just straight up lost to Tokugawa, so I don't want to say that his lack of success is a matter of luck.

Of the group, I feel the most confident about Victoria, Gilgamesh, and Willem, all of whom seem like they have a win in them. Isabella always seems to have moments of brilliance, but her singular focus on religion gives her real problems as the game progresses - I don't think it's a coincidence that she falls apart right around the time a leader's choice of tech really starts to matter (that is to say around late Renaissance/early Industrial). Peter and Louis have both made the playoffs so they don't *really* belong in this group, but they likewise can't seem to close for whatever reason.

Darius feels like he should be in this group, since he has a strong civ, good traits (he is financial, after all), good flavors (gold and growth), and has that nice 2 point differential between wonder and unit preferences. However, he has a high peace weight, can't declare war at pleased, and isn't religiously motivated enough to start an opportunistic war against a weak enemy. It's frustrating, but he just doesn't seem to have the killer instinct the top leaders all enjoy.

[Fluffball]

• a value-added statistic for traits - in other words a way of attempting to filter out the benefit a poor trait gets from the leaders who have it paired with a strong trait
  

Expanding on this, the trait usefulness are heavily biased by what leader has them. For example according to the stats, industrious is one of the worst in the game, however I think it says more about the fact that idiots like FDR and Bismark have it than anything else. One thing that went largely unnoticed in the streams is that Stalin was KILLIN' it with wonders this season -- in every single game he played.

I view traits from the perspective that the AI is going to do everything: they will research every tech, given enough time they will build every building, they will eventually build every wonder still available, etc. Industrious saved Stalin HUGE amounts of time in that regard; getting Stonehenge meant he didn't waste the hammers building monuments in every city, he didn't waste the hammers losing SH to civs that started it earlier... Getting the Oracle for an AI, while not game changing like with a human still means they have gained a substantial turn advantage in research.

While perhaps the deciding factor in every game can be luck, the rest is just a big blob of "do everything", and those little pluses here and there inch one civ's blob of "do everything" a little closer to the finish. Mansa's famous do-everything lets him research faster, and spends less time not doing-everything.

Edit: The industrious forge bonus also helps push forward another do-everything bonus in the extra hammers... compared to bonus drydocks from aggressive. (And the happiness is great for them too.)

Now protective and aggressive contribute nothing to the do-everything. AI wars are decided by the biggest blob of doing-everything, not individual units. On top of that, deity AIs build ungodly stacks of seige which nearly completely nullifies attackers and defenders. Who cares that the lead mace has +10% combat strength when the lead longbow got hit with 14 trebs?

TLDR: Industrious is a great trait in the hands of a solid AI, it's just in the hands of an unusually large number of terribad AIs.

[GreyWolf]

Victoria's willingness to go to war seems to be a bit too high for someone with a mere 4/10 unit emphasis, lack of Military tech flavor plus a high peace weight (a drag, like Axiis observed). She is prone to declaring inept, badly conducted wars on unit spammers. Still, even taking this into account, she indeed has a strange propensity to throw away the game.

Bismarck, I think, has gotten unlucky. He isn't that bad of an AI, in my experience.

Isabella won't sign Open Borders at Annoyed, and she has a default -1 to everybody else. This often means she has trouble spreading her religion through missionaries.

The worst AI is probably Sitting Bull. Protective is bad for everyone, Philosophical doesn't do much in itself, decent unit emphasis, but pacifist, so he won't really do a unit build-up to gain land. Instead his units are just a drag on his economy. Hate rod for warmongers, but unlike a Gandhi, he has no chance of out-building anyone. The only thing totem-ed archers and longbows do for him is that eliminating him takes longer. If he manages to ever win a game in a similar competition, I'll be truly shocked.

[Eauxps I. Fourgott]

So while there are a number of factors that contribute to the poor performances of high-peace-weight leaders, I think it is worth noting that one factor is that there are simply less of them. I went and counted all of the leaders of each peace weight category, and while I forgot the exact numbers, it was something along the lines of 14 good, 13 neutral, and 23 bad - and that's also considering the fact that neutral leaders have peace weights of 3 to 6, making them skew more towards bad. So really the high peace weight leaders are disadvantaged right from the start - with lower numbers, they're more likely to be in a situation where they can get teamed up on by the evils.

And so Willem is last place overall in both forms of rankings... I think one factor that often contributes to his downfall is that he tends to underexpand badly. In my experience, he'll often be slow to build more than a few cities. His opening round performance in this season was a prime example of that. I'll also always remember the game where I played on a team with the Willem AI, who then decided to do a OCC, letting Mansa have the entire long, snaky island that the two started on. I think that in general he has a tendency to do too much building upwards and not enough building outwards.

As far as leaders who have done poorly due to bad luck, I'd nominate Freddy as one of the top ones. Season 1 Game 1 had him boxed in by aggressive expansion from his neighbors to the point where despite a solid showing he couldn't recover enough to take a top spot. His other games have seen him be the target of attacks time and time again, preventing him from really getting off the ground. Darius crippled him at the beginning of the Season 1 wildcard, he got overpowered by Shaka in Season 2 (this was probably the defeat he deserved most), and Season 3 saw him working on fending off a Napoleon assault before De Gaulle made it a 2v1 and sealed his fate. I think he's still waiting to catch a break.

Anyway, thanks for another season, Sullla! While I won't deny that this season got frustrating at times (Mansa attacking my second-place pick Lincoln at the last second for no reason comes to mind, as do Kublai and Pacal's presences in the finals), it was also quite a bit of fun. Looking forward to Season 4, whenever that may be!