The Battle of 100 Freeciv AIs

This is the story of an epic battle of 100 Freeciv artificial intelligence (AI) players fighting on a very large map with 20.000 tiles!
The results, statistics and the winning AI of the game can be found here:

Statistics including player scores, population, gross national product, military units, settled area, land area, number of technologies, cities, settlers, units built, corruption, gold, technology output and  wonders were recorder each turn and presented as graphs for further analysis. Every 10 turns, the whole game screen was captured and saved so it could be replayed online here for you.

What can one learn about conquering the world, based on a simulation of 100 AI players competing on a very large game? Which tactics and strategies were used by the leading AI players at various stages during the game? The fight between Antso III Gartzes Nagusia of the Basques (golden color in score graph) and Mściwoj II of the Kashubians (green color in score graph) is particularly interesting. These two players dominated the game over a long time, while the Basques has the highest score.

After playing over 500 turns, it seems that still no single AI is capable of conquering the whole map for themselves. Perhaps Freeciv doesn’t converge towards one winning AI, but rather sustaining multiple competing nations over time?

This AI simulation is inspired by Battle Royale by thenyanmaster.

The score for each player over the 534 turns the game lasted:


Freeciv server version built from current svn (2.5.0+). Simulating this AI game took 48 hours of 100% usage on a 4,7Ghz CPU. Please post suggestions about interesting AI simulations which could be done in Freeciv. Do you have any ideas for new AI simulations using Freeciv? For example, should the next AI simulation have 500 AI players? How about simulating alliances of AIs with the republic vs. democracy vs. monarchy and so on?

So head over to the statistics results of this fight between 100 Freeciv AIs here:
Now you can try out these strategies yourself, and challenge the AI. Freeciv-web can be played online in your browser for free here.

14 thoughts on “The Battle of 100 Freeciv AIs”

  1. That’s a cool idea ! I would be interested in the raw data (the data presented on the graphs).
    Maybe running a lot of small strong-against-weak games would also be interesting (does the latter perform better, and why ?), or games with different starting conditions (do early workers make the AIs reach high scores significantly faster ?)

  2. This is enthralling. I would be interested in statistical analysis of a huge set of games with fewer players. You could take the mean lines for each civ in thousands of 7 nation games with random starts but the same players. That way you could see the most common trajectories for things like Ottoman gold or Chinese population. If you took it a step further by increasing the weight of the winning games in the averages (for the civ that won that game), then you could make an “ideal” mean line for each stat/civ combination. That could act as a guide to winning the game for these civs! Either way, less civs will make an easier to read graph, then some stat analysis might bring up some interesting talking points….

  3. That’s interessting. What were the (server) settings you used for this game?
    Last time I played, the AIs preferred a strategy of building a lot of very small cities very close to each other. Here it looks like they let their cities grow some more.

    How about adding some kind of human irrationality to some of the AIs, for example some sort of vendetta-mode where one AI tries especially to defeat another one, even if another strategy could prove more profitable.

  4. What 4.7 GHz CPU was used? How many cores does it have? Does FreeCiv support multiple cores at all? If yes, why did the simulation take avg. 5 minutes per turn; shouldn’t it be faster?

  5. If you flip a coin once, can you tell if it is a fair coin? How about twice? Of course not. The part of statistics that says how many flips you need is “acceptance sampling”. ( The same is true for AI capability.

    If you want to know the most likely to win, (all else being equal) then you need to run them against each other many times, and with all else (like geographic starting position, and who their near neighbors are) properly and sufficiently varied.

    It would make a useful paper to document how to determine proper “statistical design of experiment” to be able to tell which AI would win given only the AI names, their 5 nearest neighbors and their location on a map. It would speak both to automated experimental design, and to comparative evaluation of AI. This might make an interesting version of a Turing test as well. Can you determine whether one AI is statistically different in competitive performance than the remaining AI’s. Fun

  6. If you are interested, I made a video out of the full resolution maps (

    The video is made at 1 fps, which gives a little bit less than one minute of video.

    It seems that YouTube allows a maximum resolution of 2160p. If you like I can upload the full resolution video somewhere else. It is only about 16MB.

    In the previous comment I posted the same video, put did upload with a different YouTube account. Sorry for that!

  7. Perhaps the AIs were designed not to totally demolish a player, thus leaving openings for other AI to continue. Other possibilities include… No AI is good enough to take over the whole map by itself? If all the AIs were designed to “give the player a fun game”, then they’re probably doing exactly as expected. I’m not sure what this all proves, really, other than you can simulate a 100 AI player civ game on a reasonably fast computer fairly quickly.

    Which means if you want to have a 100 real player game, it’ll have a minimum lower bound of 48 hours, and max of “why did we try this again?” hours.

  8. “After playing over 500 turns, it seems that still no single AI is capable of conquering the whole map for themselves”

    I think more interesting is that if you look at the graph that shows the scores, you see that no single AI was destroyed …
    Really? 100 players in a game and after 500 turns none of this was blown-up?
    Seems to me that the game set-up was designed in a way in that no one can ultimately win or loose.

    So, as interesting as this game setting is, the results do not allow reliable conclusions. At least not without in-depth view of the AI settings.

  9. It would be nice to start a challenge where developers around the world are required to implement their own artificial intelligence in order to conquer the world.

  10. Hello! Cool simulation. But too quickly finished. It took you 48 hours? It seems your CPU is just too slow to simulate 100 civilizations.

    Maybe you should take only 40-50 civilizations on a smaller (1000×1000) map, but make a much longer simulate (like, 2000 turns? or simulate till one player left?). That would be much more interesting than simulating only 500 turns.

  11. Why didn’t you run it longer than 48 hours? Eventually one civ will hit the space race and actually win, right?

Leave a Reply

Your email address will not be published. Required fields are marked *