Quick civ analysis

[fatherbushido]

Here are some results from tests done one year ago.

#Protocol

- 50 games 1v1 on mainland with a set of predefined (random) seeds, for each couple of civs, both players positions test

- players are AI

- tested on a fork of zeroad

#Win-Loss results

For each couple of civs, we compute the difference between number of wins and number of losses.

In that context:

- rome, brit, gaul are far ahead

- kush, mace, pers are far behind

There is perhaps an high civ unbalance or it is related to the way the AI use them.

 

 

#Trading score

There is the potential bias that it's easier to trade when wining the game.

Pers, despite losing a lot of game, shows high trade income. That reflects there high trading bonus.

#Trained support citizen

 

#Trained infantry cavalry ratio

 

From AI point of view :

- Brit, sele, ptol look like cavalry civ

- Athen, spart look like infantry civ

 

#Siege and champions

Elephants are counted as champions.

It's strange that mace that is expected to be a siege civ don't produce a lot of siege.

Kush and maur don't have siege units. One can notice that maur gained some by capturing an enemy structure.

AI may miss something or the civ is too late at phase 3.

 

Enjoy!

PS: That's a quick approach, there was some extreme games messing the means.

 

Edited February 26, 2021 by fatherbushido
That forum has issue with image insertion

[fatherbushido]

Do you think showing elephants in those stats is relevant?

[ffm2]

I dont know the balance of your fork, but i'd compare it to the 0ad a23 balance, I'd say it only reflects the strategies the bot uses.
Persians can be very strong in the late game using cav and the according hero. This of course requires the bot to build his game accordingly from the start towards that strategy.
Mace had the advantage to get siege fast in their workshops.
Your analysis is good, but I think this should be generated from real games from players over 1400.
On lichess for example all games are recorded and are open available, I don't see a problem why 0ad shouldn't collect at least some data, like the civ.
With this https://trac.wildfiregames.com/ticket/4376 you could generate it out of the logs.

[fatherbushido]

Thanks for your interesting input! You are absolutely right.

We wanted to detect ai flaws and civ flaws through those 3900 games.

The ticket you point out would be a good complement! Thanks for pointing it.