Curious about Dominions? how the AI thinks? which AI is best/worst? - Page 2

JimMorrison · #11 January 9th, 2009, 07:53 PM

Quote:

Originally Posted by Gandalf Parker

It occurred to me that it would be helpful if someone wanted to provide data for mp to go with this. Rather than impressions and opinions maybe someone administratively skilled could break down the winners list from the sticky in the mp forum? Just totals for each nation. It wouldnt provide the 1st out info but it would provide winner stats

Actually, many of those victory posts contain links to the original game threads - so many of those, will contain posts stating that one nation or another was defeated, and thanking someone for playing. So theoretically, with a few long hours of effort, some data could be pulled. The problem being how do you quantify it, if out of say 10 nations, 1 wins, 5 are defeated outright, and 4 others are alive at the end of the game?

Gandalf Parker · #12 January 9th, 2009, 08:06 PM

The equip changing thing may be a fairly new change. I dont remember seeing it before either. Its been a couple patches since I last looked at a game log much.

Gandalf Parker · #13 January 9th, 2009, 08:11 PM

Quote:

Originally Posted by JimMorrison

Actually, many of those victory posts contain links to the original game threads - so many of those, will contain posts stating that one nation or another was defeated, and thanking someone for playing. So theoretically, with a few long hours of effort, some data could be pulled. The problem being how do you quantify it, if out of say 10 nations, 1 wins, 5 are defeated outright, and 4 others are alive at the end of the game?

Quite right. Its definitely not going to be as clinical a result as the AI tests are.

Ive gathered some data from the games that happen on my server. In fact its one of the reasons I started doing it.

But it doesnt have to be great data. Im HOPING that it tells us nothing at all. That would be best. But if the same nations seem to fall out of the game in the first wave, and the same ones appear to top the winner list, all with different people playing them... then that would be data worth noting.

People have often said "everyone knows" facts about weak and strong nations. When Ive glanced thru things I was never able to spot a correlation.

Gandalf Parker · #14 January 10th, 2009, 03:01 PM

Its kindof fun watching the nations rise and fall.
Im thinking I could make a killing by giving odds and taking bets as each game progresses.

VedalkenBear · #15 January 10th, 2009, 04:20 PM

GP, I'm currently taking a course this semester that will require me to learn how to do all kinds of variance analysis (basically, correlation and covariance).

If you can get me the data, I can perform this analysis and comment usefully on the outcome.

In addition to simple placement, I (personally) am interested in the strength of a nation related to the distance it starts from every other nation. This might be a good measure to determine the shift in strength of a nation from early/mid/late.

As an example, many people claim that certain nations are vulnerable to rushes. Since not all nations can perform such rushes, an analysis comparing the 'rank' of a vulnerable nation to the distance it starts away from the rush nations might reveal whether there is, indeed, something to examine.

Gandalf Parker · #16 January 10th, 2009, 05:35 PM

The options for finding that out is:
peek into the game (which the methods corrupt the data)
Pre-set their locations on a map (which limits the data)
Or set the map size and collect lots of data to even out the numbers.

Im working with the last one though it takes the longest. At the moment all games are "medium" games by Dom3 standards. Im using a --randmap 15 on the game creation which generates a random map of 15 provinces per player. Later I will try the same with --randmap 10 (small game), and --randmap 20 (large game). And then I would be really curious to try it with something like --randmap 50 just to see the difference.

There is also another factor which has to be figured in. The AI uses type 1, type 2, or type 3 AI. Standard, Defensive, and Aggressive. I have no way to set those so every game, every nation, will receive a random AI.

So the end ratings would be 9? for each nation?
Early Ulm:
Small: std, def, agg
Medium: std, def, agg
Large: std, def, agg

I guess that each one could be just a number reflecting its avg rating. In Early Age there are 24 nations. Lets say 24 means usually first out of the game and 1 means usually winner. so if it was
Early Ulm:
....s.d.a
Sm:3:5:1
Md:6:9:3
Lg:10:13:8

then we would be able to show (at last as far as the AI handles Ea Ulm) that its game balance tends to be that it does well on quick-game/small-maps. Hopefully another nation shows exactly the opposite. (my hopes for that was on Arcos but Im beginning to show that Arcos has a problem which I will bringing to Johans attention soon)

BTW for those who prefer more than numbers. Im working on charts. And Im working on a "Watcher Nation" which will allow a Gods Eyes view of the game with many super stealthy scouts (flyers, maybe teleporters). It will be a non-playing nations which will allow "tourists" into the game. Not THESE games since it would violate the data but I will make some mega-huge AI game (possibly with players) available for spectating.

VedalkenBear · #17 January 10th, 2009, 11:15 PM

GP: Yes, something like that can be done. My question is if it's possible to inspect various data to find patterns; e.g., is it possible to determine how far apart all the nations start?

Gandalf Parker · #18 January 10th, 2009, 11:42 PM

Not easily Im afraid. All we can go by is the outside info of small, medium, large map.

If I could view the map I would be more interested in who the neighbor is than how close it is. If Undead Ermor starts near Marignon then its weaker than if it starts near Ulm.

Gandalf Parker · #19 January 11th, 2009, 10:42 AM

Wow. The latest Mid_AI test has run all night. Its too bad that the SuperStats stops recording at turn 200. That was fine when there were few nations but now thats definetly not the "game as good as over" point.

At turn 200 it was:

Code:

Stats turn 200 for game 'Mid_AIs'
Ermor 		 provinces 37 	 inc 690
Man 		 provinces 47 	 inc 2464
Ulm 		 provinces 4 	 inc 76
Marignon 	 provinces 52 	 inc 2369
Machaka 	 provinces 30 	 inc 792
Agartha 	 provinces 54 	 inc 2063
Caelum 		 provinces 7 	 inc 605
Shinuyama 	 provinces 1 	 inc 3
Ashdod 		 provinces 4 	 inc 23
R'lyeh 		 provinces 65 	 inc 2617

And I might have bet on Marignon or Rlyeh. Maybe Agartha.

When I went to bed the provinces were at
Agartha: 116
Machaka: 81
R'lyeh: 54
Man: 50
I would have bet that in a couple of hours Agartha was going to win with Rlyeh coming in second (just because the water nation tends to be the last taken out)

I wake up and its
Man: 106
Machaka: 97
Agartha: 95
Rlyeh: 0

Man in the lead.

Now 20 minutes later its Machaka in the lead 112 to Mans 99 and Agarthas 88.

Another interesting point is that the SuperStats shows
Man (ai 2 type 3)
Machaka (ai 2 type 2)
Agartha (ai 2 type 1)
Man is playing with aggressive AI, Machaka with defensive AI, and Agartha is with standard (middle of the road) AI.

VedalkenBear · #20 January 11th, 2009, 12:45 PM

Quote:

Originally Posted by Gandalf Parker

Not easily Im afraid. All we can go by is the outside info of small, medium, large map.

If I could view the map I would be more interested in who the neighbor is than how close it is. If Undead Ermor starts near Marignon then its weaker than if it starts near Ulm.

Mm, that's a shame. The closeness of the neighbors will show the 'who' as well as 'what'. For example, if you're examining the probability that Ermor will win in the MA or LA, you can run a correlation analysis on the winning percentage (or perhaps 'xth to die') on their original distance from Marignon. If the analysis returns that there is a high positive correlation with distance, that would mean the the closer Ermor starts to Marignon, the worse they do.

Following that logic, if, say, MA Ulm consistently has a positive correlation with initial distance from most/all 'rush' nations, one inference you could draw would be that MA Ulm is highly vulnerable to rushes.

Now, I don't know much about the various switches involved, but... is it possible to manually set up a game, record the initial positions, and then set all the players to AIs, record what AI they receive, and use this utility to get the information?