AI League 2.0

[Genghis the Barbarian]

That is the first time I have heard about an AI launching with incomplete Casing.

Happened to me once in one of my S3 AHs

 

[Thrasybulos]

I definitely remember watching a game where a part got sabotaged, but IIRC Sulla had not been paying attention to this detail and dismissed it as an unexplained oddity.

In the S8 championship I noticed Gandhi getting his spaceship parts blown up a lot especially after going nuclear and I think that really really helped Lizzy. Without that I think he might have won

So it would seem the apparent lack of sabotaging in AI Survivor owes either to Sullla being oblivious to it, or to my faulty memory, or both.

Either way, the settled Great Spies get innocented.

I must say they're a bit of a puzzle to me: logic would dictate they do have an impact on the game.
Having the AIs with their espionnage slider locked at the max value (20%) for the whole early game and part of the mid-game, wasting a ton of EPs on the Observer civ, should mean a slower game pace and thus later finishing date, more AIs hopelessly crushing their early eco, and less inter-AI espionnage action later on... and yet I can't say any of that has been noticeable.
Which is good news for AI Survivor, btw. But still puzzling.

 

[Thrasybulos]

Series 11 Results

Spoiler Results :

Well... third time's the charm, apparently.
Once again, Ramesses stayed at the top the whole time, but this time he was able to turn that into a win!

The difference this series was that the high peaceweights performed a lot better overall, so Ramesses usually had at least one potential ally in Pool 1, and that was enough. The fact he also was the only AI to always get a win from the strongest start on that map surely helped as well...

As an illustration of the high peaceweights doing better, on round 10, Pool 3 was 100% high peaceweights (Victoria, Gandhi, Mansa, Darius, Asoka, Pericles).

Mansa and Hatty both had a strong start, but when Hatty got stuck in low peaceweights groups, it looked like Mansa would finally catch up to her.
And then, in round 8, this happened:

I suppose a couple of those were lucky wins, but still...
That propelled Hatty to Pool 1, which benefitted... Ramesses.

Huyna Capac did OK, but by his standards, he clearly underperformed.
That was also the case of quite a few other low peaceweight leaders: Saladin recovered in the latter half of the series, but Hannibal or Kublai for instance had a really bad time overall.
And while he's not a top AI, Mehmed in particular performed disastrously here.

With the high peaceweight doing better, the lower pools were for once dominated by the low peaceweights.
I guess it's no wonder that on round 10 for instance, the only high peaceweights stuck in that hostile field were Frederick and Sitting Bull...

At this point, I don't think the following is likely to change:
Best 3 low peaceweights leaders: HC, Willem, Justinian.
Best 3 high peaceweight leaders: Ramesses, Hatty, Mansa.
Well, turns out Round 4 had that exact matchup.

... and that didn't go too well for the low peaceweights, for once.

Spoiler Awards :

No new record this time, but Frederick almost appeared twice in the Calimero awards (he gets the 1st and 4th spots)!

Spoiler Elo Ratings :

Ramesses is the big winner, overtaking Pacal for the 4th spot, and joining the 1900+ club.
The biggest losers are Frederick (which changes... nothing as far as his ranking is concerned) and Mehmed who had a really bad time this series.
Also note that the Darius vs Zara contest is still unresolved, both still only separated by a single point.

Spoiler Map Data :

Spoiler Pool 1 (S5 Game 7) :

Spoiler Pool 2 (S4 Playoff 2) :

Spoiler Pool 3 (S8 Playoff 2) :

Spoiler Pool 4 (S3 Game 7) :

Spoiler Pool 5 (S5 Game 2) :

Spoiler Previous Results :

Spoiler Pool 6 (S3 Game 2) :

Spoiler Previous Results :

Spoiler Pool 7 (S1 Game 8) :

Spoiler Previous Results :

Spoiler Pool 8 (S6 Game 6) :

Spoiler Previous Results :

 

[AcaMetis]

No new record this time, but Frederick almost appeared twice in the Calimero awards (he gets the 1st and 4th spots)!

The biggest losers are Frederick (which changes... nothing as far as his ranking is concerned)

The saga of the man who keeps on digging continues .

 

[Thrasybulos]

Hammurabi's spaceship was due to land in 1949, but he reached the Domination threshold on the same turn.
When that happens, the Domination victory condition is always triggered first.

 

[Genghis the Barbarian]

View attachment 749299

Hammurabi's spaceship was due to land in 1949, but he reached the Domination threshold on the same turn.
When that happens, the Domination victory condition is always triggered first.

IIRC it's by alphabetical order, so Culture > Diplo > Dom > Space. If there are ties for the same victory IIRC it's a coinflip

 

[Thrasybulos]

Could be as simple as that indeed.
The coinflip for Space has been confirmed and documented earlier.

I'm certainly encountering a fair few examples of the Domination vs Culture race, but it's harder to pinpoint the exact date each is due when running aiplay, so those go unnoticed.
There might be an instance in my League 1.0 attempt: since I was running those games manually, I should have documented it. Would need to check it out.

 

[Thrasybulos]

Series 12 Results

Spoiler Results :

Back to business as usual, with Huyna Capac at the top?
With 5 more wins than any of his challengers, sure looks like it, but... instead of him being dominant, it was actually him being the most consistant.

He started super strong, netting 5 wins in round 1, with Ramesses and Suleiman netting 4 each.
But after that, Justinian was the best performer, and it looked like he would at last get a Series win. And then, as it happened in a previous Series... he just stopped winning.
But it was Gilgamesh, not HC, who took up the slack and started to look like he was headed for the final win. Then he, too, faltered.
The final rounds saw Shaka emerge as a real beast: with just one more round at a similar performance level, and he'd have been the winner.
But thoughout those three periods which saw a different dominant leader, HC was consistantly the second or third best-performer... which was enough for him to finish ahead of everyone else.

The "back to usual business" aspect was definitely there, tough, as far as pool composition went: the top 3 pools were dominated by the low peaceweights.
Ramesses did well, as usual, but in such a hostile field, had no hope this time.
Hatty also started well (slower start, tough, as her initial round went poorly), but got crushed in the top pools.
Mansa... had a harder time.
But both he and Hatty had a stellar round 9: four wins each! Unfortunately, that meant Hatty "overshot": she ended up in Pool 1 for Round 10, while Ramesses and Mansa landed in Pool 2. So she got slaughtered, while Mansa and Ramesses couldn't work a miracle. Things might have turned out differently if she'd been able to join up with them.
Then again, Pool 2 had a slight map problem... (see below)

Gandhi fared rather poorly, and unfortunately for him, he wasn't able to pull off a last-minute come back.
Sury, on the other hand, did pull one off: can't say he had actually "fared poorly", but his performance had been mediocre. But he pulled 3 wins in Round 7 and then again in Round 10, and that saved his tournament.

Frederick doesn't end up last... only next to last!
Instead, it's the other German leader, Bismarck, who closes the ranks. Bad Series for Germany!
That said, for most of the Series, it was actually Boudica who was firmly sat in the last place, with a single win scored for a long time...

Oh, and look, for once Qin and Mao are respectively sitting where I'd have placed them before I started running this league!

Spoiler Awards :

Shaka beating his own kill record came at no surprise.
But Mansa didn't feel like he was having a Gandhi day, and yet... most eliminations, but fastest Cultural and Space victories.

Spoiler Elo Ratings :

At the top, Huyna Capac, Justinian, and Ramesses had a good result.
But the best performances, and it shows here, were from Gilgamesh and Shaka. The latter is in hot pursuit of Napoleon, who underperformed somewhat. Did I speak too soon when I said it was clear Napoleon was the best of the crazies?
Bismarck and Boudica register the biggest losses here, closely followed by Peter.

Darius seems to have pulled a tiny bit ahead of Zara, but that owes entirely to the Round 10 result: they were still only one point apart until then!

Spoiler Map Data :

Spoiler Pool 1 (S8 Game 3) :

Surprisingly fairly balanced.
Position 3 (Freddy's in AI Survivor's game) does prove the strongest, as had been correctly identified at the time. Ironically, though, I seem to remember Sullla pointing out that if HC had started there, that would have been a slam dunk... while one of the reasons he wasn't super dominant in this Series is that his track record from that spot is rather poor: 3 wins only out of 9 games.

Spoiler Pool 2 (S8 Game 5) :

Alright, we might have here a contender for the title of most unbalanced map!
And cut Darius some slack, will you? In the AI Survivor game, he started in position 2...

Spoiler Pool 3 (S5 Playoff 2) :

Spoiler Pool 4 (S2 Game 3) :

Strong effort, but the Pool 2 map beats it.

Spoiler Pool 5 (S4 Game 4) :

Spoiler Previous Results :

Spoiler Pool 6 (S7 Game 8) :

Spoiler Previous Results :

Spoiler Pool 7 (S3 Game 4) :

Spoiler Previous Results :

Very different result for Position 4: it was the second best spot here, above average, and it's the worst spot in the new result set!
I believe part of the explanation lies with Position 7: their fortunes seem linked. If the AI in position 7 does well (which happened often in the new result set, whether "doing well" results in winning or not), then the AI in position 4 will not. And vice-versa.

Spoiler Pool 8 (S8 Game 4) :

Spoiler Previous Results :

 

[Genghis the Barbarian]

I’m kinda interested in seeing series 13-15 be played on each of the three mirrored championship maps that have been used

 

[Thrasybulos]

The plan is that Series 15 will be the last one (since I always have more ideas than time/availability, no worry, there's more AI stuff coming afterwards ).

That will allow me to have gone through all standard maps, with two "spots" available in Series 15.
The S4-S7 Championship map, once the errors are fixed, is perfectly mirrored so not a lot of information to gather about the map from running it (with the exception, maaaybe, of the influence of the start position regarding Culture mode?).
So I'm planning on having the S1-S3 and S8 championship maps as the "guest maps" for Series 15.

By the way, have you had a look at this discussion over at Realms Beyond? If Sullla's planning on using the S8 map again for S9, you might want to reuse those edits?

 

[Fippy]

I didn't like that map..unfortunately.
Too rich and somewhat sheltered branches, could warmongers ever catch the likes of Mansa or HC here?
Mansa maybe..cos of peaceweight, but i imagine HC would have the time of his life here

 

[Thrasybulos]

i imagine HC would have the time of his life here

Except when he has Shaka as a neighbour, possibly?

 

[Fippy]

Okay but how often are (real) warmongers in CS games?
Usually the best we can hope for are AIs like Giggles, Kublai, Louis etc.

I didn't verify now, but think neither Shaka - Nappy - Monty - Alex - Ragnar - Genghis - Boudi or Toku ever made it that far.

 

[Thrasybulos]

That's a fair point.

 

[Genghis the Barbarian]

The plan is that Series 15 will be the last one (since I always have more ideas than time/availability, no worry, there's more AI stuff coming afterwards ).

Makes sense. If possible I'd be willing to help out on some (in particular I'm interested in "unrestricted leaders" setups)

That will allow me to have gone through all standard maps, with two "spots" available in Series 15.
The S4-S7 Championship map, once the errors are fixed, is perfectly mirrored so not a lot of information to gather about the map from running it (with the exception, maaaybe, of the influence of the start position regarding Culture mode?).
So I'm planning on having the S1-S3 and S8 championship maps as the "guest maps" for Series 15.

I think there's quite a bit more to glean from mirrored maps than it may seem. I've certainly noticed that just because a map is mirrored, by no means is it balanced. For example, when running or reading AHs from the S1-3 map (henceforth "Circle"), one observation I noticed is that 1) the map seems to be pretty bad for cultural leaders and 2) the map is also bad for Creative leaders. On the latter point, I observed (at least from Kublai Khan in S3 Champ AHs) that his Creative trait was causing him to settle fewer cities (because the AI often don't settle within cultural borders even if they could) and thus he often had 6 cities when a Caesar or even a Pacal had 10, leading to a high First To Die rate. Sury did very poorly in S1 Champ AHs. Both of Mansa's sets have been unmitigated disasters, and Lizzy was not great in her set. Running a series with the Circle map could give insight into which leaders benefit more (i.e. militaristic leaders, Imperialistic, Fishing starting leaders perhaps?).

The circle map also seems to have given full on cultural rolls - my S3 AHs had games where the winner nearly blew with late culture runs - which probably explains why culture leaders struggle because being in culture mode on T0 is horrible for a variety of reasons.

Conversely, we could answer for once and for all whether the S4-7 "Rectangle" map favored Creative leaders a lot more, or if the "Star" map favors peaceful or religious leaders. Lots of questions. How much does neighbor or peaceweight situation matter (I'd say a ton for Circle, a good amount for Rectangle, not nearly as much for Star)? Are the resources spread out in a way that benefits certain starting techs more? Culture mode (as you said)?

By the way, have you had a look at this discussion over at Realms Beyond? If Sullla's planning on using the S8 map again for S9, you might want to reuse those edits?

Skimmed through a bit. I do think some improvements are needed - at least put in fewer resources. My idea is to also put starting positions much closer to the center. Would lead to interesting decisions - try to claim the center and intrude into one leader's "tentacle", or develop peacefully in your own tentacle? Maybe you could play around with it for the Star map if you run an entire series on it.

 

[Thrasybulos]

I've certainly noticed that just because a map is mirrored, by no means is it balanced.

That's my intuition as well.
When I started doing those experiments a few years ago, it was suggested to me I should use mirrored maps.
But I thought such maps risked being biased towards certain leaders, so I opted instead to see how leaders would fare when exposed to a great variety of starts.
And as a bonus, it also provided an assessment of the various starts.

And while you're correct there'd always be something to learn on any map, I can't be running after every goal post out there.
So I'm limiting my assessment of maps to comparing how strong or weak each start is.
In that context, a fully mirrored map offers little interest.

 

[Thrasybulos]

Would lead to interesting decisions - try to claim the center and intrude into one leader's "tentacle", or develop peacefully in your own tentacle?

Apart from the objections that Fippy has raised, and which were confirmed by @Keler or other tests I ran, that would be my second objection: those are interesting decisions for a human player.

But we're evaluating the AIs, and we know the AI code doesn't take such high level strategic aspects into consideration.
So those "decisions" are an illusion.

Not throwing the stone there, because I went there too: some years ago, I tried exactly that. Made maps which I believed would provide some strategic depth... and watched the AIs play and completely ignore all those aspects.
And then I had my "duh!" moment.

 

[Thrasybulos]

Wow, that came very close to being a Conquest victory!

 

[Thrasybulos]

So, the various projects I have in mind, post-League.

1- Horatius Trophy

This would be a spin-off of the Horatius League, with each game recorded as a video.
Instead of a Football World Cup inspired format (pool phase followed by a knockout phase), this would be a single-elimination tournament throughout.
Take 12 seeded leaders which are qualified directly for round 2.
The remaining 40 play the 1st round: 20 go out, 20 make it to round 2.
20 + 12 = 32, things can now proceed in a standard fashion henceforth.

That thing is actually ready to go at any moment: I worked on balancing the map which would be used back in September, and that's basically all that was needed. I was planning on an early October start, but RL got in the way, and then I went "cold feet".
Even though that tournament would feature far fewer games than the almost 400 games which made up a single Horatius League "tournament"... we're still talking 51 games here.
That's a pretty big commitment. I would owe it to viewers to see the thing through, but recording 51 videos is pretty time-consuming, and if I get only a handful of viewers (pretty likely, actually)... is that worth it?
I could shorten up things a lot by recording aiplay games of the first rounds, reserving the full manual games for the final rounds... but that would probably lower the interest for a project which would have a very niche appeal to start with.

So I dunno...
That project isn't shelved, but on hold for now.

2- AI "schuffle" (shuffled scuffle)

I've already hinted at it: it would be a continuation of the AI League, but with the Swiss Format dropped.
Take four 6-player maps, four 7-player maps, and play 52 rounds with random match-ups, having as the only constraint that each leader plays from each starting position exactly once.

This could be started pretty fast: I've already designed the database structure changes, I've written and tested the code for game generation. I would just need to update my helper application with those changes and rework the interface a bit. One busy week-end should cover it.

The only aspect I was still puzzling about was: should I keep using the AI Survivor maps, or move on to a new set of maps?
And I believe the answer is... both?
A first run with the AI Survivor maps would allow me to compare the results with the League's results, and thus assess the impact of the League format.
And a second run using different, more balanced maps, would allow to assess the impact of the imbalance of AI Survivor maps.

3- AI Survivor 2.0, King of the Hill edition

Sullla's base idea was "an eliminated AI is out", and he worked up a tournament format from there.
I had some discussion with @Keler last year about it, because I believe a better base idea would be "only winning AIs move on".
After thinking about it some more, I believe I've finally come up with a format I'm satisfied with.

Start with 49 AIs, not the full 52 complement.
Play seven 7-player games. The 7 winners qualify for the playoffs.
Play six 7-player games with the remaining 42 AIs. The 6 winners also qualify for the playoffs.
Play six 6-player games with the remaining 36 AIs. The 6 winners join the other into the playoffs.
And that concludes the opening round.

We have 19 leaders qualified for the playoffs.
We add two of the three AIs who didn't play the opening round: that makes 21.
Play three 7-player games. The 3 winners make the championship.
With the remaning 18, play three 6-player games. The 3 winners join the championship, the others are out.
We now have 6 AIs qualified for the championship game. We add the last AI who hasn't played yet.

The championship game is played with 7 players.
We conclude the tournament with a 6-player game to determine the tournament's runner-up.

The 3 AIs left out from the opening round games would be:
- The current reigning champion, who'd be directly qualified for the championship game as the defending champion.
- The previous tournament runner-up, qualified for the playoffs.
- As for the last one, also qualified for the playoffs, I'd say the AI who's played the most turns in the previous tournament.

There'd be little point running this with aiplay (the League and its follow-ups are better suited for data-collecting), so this is another one which would in video format.
It be easy to get started (no need for dedicated tools)... except for two things.
First, I would need maps. And contrary to AI Survivor, I would like somewhat balanced maps. That would require quite some work: I would aim at having 72 maps (seven 7-p maps and seven 6-p maps per climate setting, plus a custom map for the championship and runner-up games each). Each would require, I reckon, about 50 aiplay games to get to a satisfying result.
So yeah... that would take some time.

And then there's the tiny issue that, better format or not... that's still basically AI Survivor.
So is there room for a second AI Survivor?
I believe I could run this between AI Survivor seasons, but would there be appetite for it?

4- AI "world conquest"
This is my latest idea, so the paint isn't dry on that one.
It would be something different.

First, create 52 12x12 "tiles" and lay them in an offset pattern to build a toroidal "world map".

At the start of a game ("campaign"), randomly assign each leader to a tile.
Then the game proceeds as follows:
- Pick a tile. Generate a game map from that tile and the surrounding tiles.
- aiplay the game.
- If the winner is the leader from the center tile (the picked tile), nothing happens (and believe me, that bothers me a lot, so it's still the source of some serious brain churning).
- If the winner is another leader, it gets control of that tile (ie, it expands its territory to now include that tile as well.
Two sub-possibilities: if the central AI got eliminated in the course of the game, then the winner gets control of the tile with his civ. But if the central AI survived, then the winner gets controls of the tile, but keeps the civ of the eliminated leader.

For example, the tile with Pacal of the Maya is picked.

The game is played, and Boudica of the Celts wins, but Pacal survives.
The tile formally owned by Pacal of the Maya is now owned by Boudica of the Maya. And it borders a tile owned by Boudica of the Celts.

And we keep going that way until a leader or a civ reaches a certain threshold (14 tiles?).

Now, the idea would be to add a meta game on top. Something like players bet on the outcome of each game to earn "reputation", and they would spend that reputation to gain "influence" with the leaders and civs.
When the game ends, the player with the most influence is the winner.

So, that one is still pretty much in the design phase... Available Q2 2027 ?

in particular I'm interested in "unrestricted leaders" setups

That got me thinking...
And I believe I could use the framework for the "AI Schuffle" for that.

Get four 7-p maps and a 6-p map. Ideally they should be somewhat balanced (by "somewhat balanced", I mean that if you played 70 games on such a map, you'd get 5-15 wins from each position. Not 40 wins from the same start and 0 from another).
That gives you 34 starting positions.

And from there, you can generate a set of games where each leader gets to play every civ once, from each available start.

Now, that set of games would be 34*52*5 = 8,840 games.
Basically the same magnitude as the whole League (15 League series is 7,800 games).

That's a lot... But should be doable.

 

[Keler]

Thanks for adding Leader Elo Ratings History, so at the end you can make a graph showing how stabilized your ranking is over the tournaments.

If anyone wants to compare championship maps, I remember making a 6 player only tournament format which was almost same as Sulla's survivor format.

8 openning round games of 6 player =48 leaders
9th openning round game a.k.a wildcard= last 4 leaders + 2 best end game scoring third finishers of the first 8 games having a second chance.
So to sum it up, last 4 leaders starting from 9th openning round game is just tiny bit unlucky.
Then we have our typical 3 play offs and 1 championship game.
All games can use championship or any symmetrical map.

So from there same openning round leaders can be placed in all three maps, then a ranking to compare maps and which leaders perform better in which map. Maybe I can bother doing that..

Restricted leaders sound like a chaos and I don't have any opinion about it. Maybe specifically creative trait assessment, Hammurabi and Gilgamesh traits can be swapped, same games of a already done test played again to see how they perform again, if possible to do without needing to edit world builder files again.

AI Surviver 2.0, Nice but why not start with 52 leaders anyway?


-52 Remaining leaders. Play four 7-player, four 6-player games , 8 winners
-44 Remaining leaders. Play two 7-player, five 6-player games, 7 winners
-37 Remaining leaders, Play one 7-player, five 6-player games, 6 winners

Now we have
21 winners ready to play three 7-player play off games, 3 finalists
18 remaining winners ready to play three 6-player play off games, 3 finalists

a 6 player championship, only winners advanced. But did you really want a 7 player championship game?

Whatever you do, or just none of these but continuation of the current league I am happy to follow you up.