2.7 AI-only Game Statistics (update: now with summary list of games)

[pineappledan]

Which is pretty close! In academia the number usually used for minimum required samples before applying C.L.T is 30, and Venice only has 28 games, but it still goes to show that the binomial assumption isn't terrible. Ideally of course all the civs have well over 30 games but that's hard to do when my computers are running 4UC games now

Woo!

 

[azum4roll]

AI might not know how to lock growth

They definitely do, and they build public works when unhappy too.

 

[CppMaster]

They definitely do, and they build public works when unhappy too.

Good to know

 

[Zuizgond]

We need a 4UC version of this, I'd be interested in seeing the difference.

I'm pretty sure Iroquois would perform even better, thanks to prowlers.

 

[CrazyG]

I would be interested in seeing which social policy trees and ideologies correlate with victory. Would it be possible to have the utility also collect this information? It might be good to record religious beliefs (or at least which civs found a religion) as well.

 

[CrazyG]

Here's a graph of the civs win rate against their score on turn 100. A few interesting observations to see:

 

[pineappledan]

The main takeaway I found from last time I compared scores to win rates is that they are basically not correlated.

 

[stii]

Score is a pretty terrible way of seeing who is winning except in the broadest terms. While it does include the important things it also includes lots of very unimportant that also give lots of points.

 

[Legen]

To confirm, is the AI still getting GAP from its difficulty bonuses? That's something that I've argued in the past about Brazil's performance months ago, and that may be affecting the results. This is also something I'm considering making a proposal about.

 

[L. Vern]

The main takeaway I found from last time I compared scores to win rates is that they are basically not correlated.

IIRC that was comparing scores at game-end time, It was possible that score at 100/200 turns could be correlated better but looks like a pretty weak correlation also

 

[CrazyG]

IIRC that was comparing scores at game-end time, It was possible that score at 100/200 turns could be correlated better but looks like a pretty weak correlation also

I calced these values for correlation between win rate and score (this is win rate, not individual games since you didn't provide that)

Turn 100: 0.27
Turn 200: 0.33
Turn 300: 0.37

I would guess that what's going on here is that having a low score is a good indicator of being a loser, but predicting the winner using score is pretty difficult. If possible, I'd track who has what social policies for more interesting exploration.

 

[KlHannibal2]

Imo probably only Inca, Iroquois and Russia can actually do it (if they don't pick another victory condition first).

And Songhai if the map has good rivers.

 

[balparmak]

This raises the question, how much does the AI consider the score when setting its rivals, or deciding its gameplan? If it heavily relies on score while score itself isn't a good metric, that's problematic and needs to change. @Recursive any thoughts?

 

[Recursive]

This raises the question, how much does the AI consider the score when setting its rivals, or deciding its gameplan? If it heavily relies on score while score itself isn't a good metric, that's problematic and needs to change. @Recursive any thoughts?

The player score? Not at all.

 

[CppMaster]

Have the score ever been adjusted? Because the time victory may be unbalanced otherwise. Or we don't care at all about the time victory? I like that in Humankind or Stellaris you win when you develop the greatest empire (by having the highest score), so the metric has to be adjusted in order to judge fairly.