After thirteen games of inter-AI warfare, the Survivor competition has finally come to a conclusion. Let's start by looking at a final ranking of all 52 leaders that took part in these matches.
I could have put this in an HTML table, but it was a lot less work to screenshot my Excel worksheet and upload it instead. I hope that the information still comes across clearly.
I wanted to rank the AI leaders based on objective criteria, rather than personal feelings about who "looked the best" or who was "the most impressive" in their opening round games. No one will ever agree on those things, and there's too much room for individual opinion. Instead, I simply listed the AIs based on how far they made it in the competition, ranking dead AIs at the same level by how long they managed to survive. While it's an imperfect system, it does reflect the fact that someone like Hannibal or Victoria played a stronger game than someone like Ramesses or De Gaulle. I think it works well enough. For the rare leaders who survived without advancing, highest score at the end of the game was used to break ties. Note that we had an actual tie for spot number 34 on the list, Gilgamesh and Joao were both eliminated on Turn 257 in their respective opening round games. For that matter, only 8 of the 52 leaders managed to avoid death and survive to the finish. Hunger Games indeed.
Here are a couple of random tidbits from the overall rankings.
* France had a rough tournament. None of the three French leaders made it out of the opening round, and the Louis / De Gaulle pair took two of the worst three spots. This was genuinely surprising, as we rate France as one of the best civilizations in the game for Multiplayer purposes. Apparently this does not hold true in the hands of the Deity AI leaders.
* Amongst other civs who had multiple leaders competing, Egypt was another disaster with Ramesses coming in 51st place and Hatshepsut finishing in 40th. (Again, Egypt is another excellent civilization, one of the best in Civ4! Strange result.) At the other end of the spectrum, there were four civilizations that put on a great showing. Greece and the Celts both put their pair of leaders into the playoff matches, something that none of us were expecting, given that neither one is seen as a good Multiplayer civ. China did even better, with Qin and Mao finishing 11th and 7th places, respectively. Mao in particular narrowly missed out on a spot in the championship match, coming up short on score points in his playoff game against Elizabeth. Finally, the three Russian leaders all had good showings too, with Peter falling in the Wildcard match and the Catherine / Stalin combo making it to the playoffs. Catherine was especially impressive, and she'll be a favorite if we ever do another one of these games.
* There were some unexpected results when it came to the trait pairings that performed the best. Let me detail this further in another separate table:
This is another shorthand way of summing up the results of each trait. I've added together the finishing position of each leader with each trait; a lower score here is better, since it means that the leader with that trait made it further in the competition. (There's also a need to work around the fact that three trait pairings are missing in Civ4, see the note provided for more explanation.) This is again somewhat arbitrary, since you could choose to weight the leader performance in a number of different ways, hopefully this will serve as a useful proxy for each trait's achievement in the competition.
The surprise of the competition was how well the leaders with the Philosophical and Imperialistic traits performed. I realized how well Imperialistic was doing about the time that we made it to the playoffs, but I was caught totally off-guard by Philosophical until I tallied up this chart. Nevertheless, Philosophical leaders had indeed been killing it for almost the whole competition. The two poor performances were from suicidally-peaceful Gandhi and "everyone hates him" Sitting Duck Bull. After those two were quickly out of the way, the other seven Philosophical leaders saw two of them make it to the Wildcard game (Frederick and Peter), and then the other five all make the playoffs, with two of them (Suleiman and Elizabeth) reaching the championship itself. I'm honestly not sure why Philosophical performed so well in these games, but it's hard to deny that track record. Imperialistic was much easier to understand, as the trait allowed AI leaders to claim a wider share of the land and then snowball off of that into dominant positions. It's clearly a powerful trait for them. Then again, if we human players got to play the game with Deity cost discounts, we'd probably make good use of settler spam too.
After that, the results were mostly predictable on this chart. It was no surprise to see familiar customers like Financial, Expansive, and Creative all performing well. These are some of the best economic traits in the game, and that holds true for the AI as well as the human. Charismatic did better than I expected, although that makes some logical sense as well. The AI struggled a lot with happiness in these games, particularly in the first 100 turns. Charismatic helps gets them through that rough period, allows them to work more tiles, etc. (I was pretty shocked at how badly the AI manages happiness when watching these games. Try spectating the BC years in debug mode sometime, it's brutal what's going inside those AI civs.)
The Spiritual trait was a disappointment in these rankings. I thought that it was one of the best ones for the AI leaders to hold, and yet the numbers on this chart didn't bear that out at all. Way too many Spiritual leaders had terrible outings, five of them finished in 40th place or lower. Hatshepsut, Isabella, Montezuma, Gandhi, and Ramesses, there's your group of shame for the Spiritual trait. Even awesome runs from Mansa Musa and Justinian weren't enough to salvage Spiritual's overall performance. Protective and Aggressive also found their way down to the bottom of the rankings. We know that these non-economic traits are bad for Multiplayer games between humans, and it turns out that they're also pretty awful in the inter-AI games as well. Aggressive couldn't get a single leader into the top ten, let alone the championship game. Finally, the Industrious trait was the runt of this litter, and it's easy to see why. Four of the worst six leaders, and ALL of the worst three leaders, together shared the Industrious trait. (Roosevelt, De Gaulle, Ramesses, and Louis - come out and take your bows!) The Industrious trait causes the AI to overbuild wonders in the early game, falling behind in the landgrab only to be swallowed up later by other, stronger AIs. It was no coincidence that Industrious also failed to get any leaders into the championship, and the first two leaders to be eliminated in the playoffs (Augustus and Stalin) also had Industrious. It was like a badge of shame in this tournament. I don't think that Industrious is a bad trait, but it's also clear that the AI has little understanding of how to use it.
I tried to see if there was any kind of correlation between AI performance and things like aggression rating, peace weight scale, and whether the AI in question would declare war at Pleased relations. Here's what aggression rating looked like when it was graphed:
I couldn't find any obvious connection here, no clear line of best fit for this data set. Peace weight and Declare at Pleased looked the same way, no clear relationship. Unfortunately, it's been too many years since I earned my minor in Statistics to remember how to do more rigorous analysis. If anyone wants to do some more work with this data set, feel free to go ahead and see what you can find.
Tracking the number of killing blows achieved by each of the AI leaders was more of a fun exercise than anything with statistical significance. This was one thing that all of the top performing AIs had in common, they were all willing to go to war and wipe out their opponents to achieve victory. Every single leader in the championship had at least two kills, for example. The playoff leaders were also heavy on the finishing blows, demonstrating how they had run over their opponents in the opening round games. Mao Zedong was the champion in this category with four kills; even Catherine only had three kills, tied in second place with Huayna Capac, Suleiman, Suryavarman, and Zara.
The real outlier in this category was Mehmed, who managed to pick up two kills (Roosevelt and Hatshepsut) despite not even making it out of the opening round. Looking back at that game in the rearview mirror, it's clear that Game Two was an extraordinary collection of AI talent. Mansa Musa was the winner of that game, and he would win another playoff game before falling in the championship. Justinian came in second place, and of course he would win the entire competition. Zara survived in third place, only to climb through the Wildcard game and another playoff game, finishing with the silver medal for the overall event. Poor Mehmed found himself in the dreaded Group of Death, I genuinely think that he'd have been a strong favorite to win some of the other groups. That's part of the fun of a random draw though! (In retrospect, Game Three and Game Five both had weak fields. Their leaders who advanced to the playoffs didn't achieve very much there.) If we do another one of these competitions, I'd expect Mehmed to have a better chance to show off his skills.
Thus we finally come to the conclusion of this event. There were lots of twists and turns along the way, and we ended up with an unexpected winner that no one would have predicted before we started. It was a real blast doing these games and running the prediction contest along the way. I've always wanted to do something along these lines, and the whole thing came together better than I ever could have imagined. At the same time, this did involve a great deal of work as well. Watching the AIs play out their games in debug mode was easy, that didn't take much effort. The hard thing was writing the reports to accompany the games, trying to describe what was going on and make it interesting to the reader. I'll be honest, when we were around Game Seven or Game Eight, I was thinking to myself, "Oh man, I've got to do fix or six more of these games still?!" So yes, it was a bit trying at times, but I'm glad that it's all finished now and I can take a break from watching these crazy AIs!
A final note. As much as we mock the Civ4 AI for its bizarre decisions and stupid gameplay, the AI programming in this game really is quite good overall. Yes, it's woefully inept compared to an expert human play. Yes, it needs major cheats on the higher difficulties to put up a challenge. That said, it still manages to play a very complex game at a solid level. The Civ4 AI on Noble is good enough to beat the overwhelming majority of human players. It knows how to research, and it knows how to defend itself from attack. You can conquer the AI empires, but they'll make you work to do it. For all the fun that I poke at the AI, I have real respect for Soren and Blake and all of the others who worked on this game's AI. They did a great job.
If you'd like to take a look at my Excel worksheets for these games, go ahead and click here. They may be of use to someone, and there's no reason not to make them available. I've also zipped together the 4000BC starting saves and the replay files for all 13 of the Survivor games. That was requested by some posters on the Realms Beyond forums, and you can download them right here. Once again, thanks for reading through these many pages of reports. I hope that you enjoyed the show, and maybe we'll do another one at some point down the road. Later.
Sullla, May 2014