AI initial road placement



  • I have been playing a lot of games vs. the AI lately, and have started registering their behavior patterns and weaknesses reasonably well. One feature I have noted: it is significantly easier to defeat the AI opponents in a 4-player game than a 3-player game. That is noticeably odd. In a 4-player game (1 human vs. 3 AI), you should theoretically win 25% of the time. It is in fact quite simple to defeat the AI 80% of the time or more. I have been trying to think through the root causes of this.

    The apparent core problem is poor AI placement of their starting roads. The AI more or less always point their initial roads toward the best available expansion location. In a 4-player game, the overwhelming majority of those hexes get taken. The result is a very high number of AI starting roads that are instantly dead (no available expansion worth pursuing). The result is devastating. A human player effectively needs 2B, 2L, 1S, 1W to expand to a third development. The AI usually needs 3B, 3L, 1S, 1W because their starting roads are non-viable very early. The cumulative benefit to the human player plays itself out very favorably.

    Some numbers here: AI opponents start the game with a total of 6 starting roads (two per AI opponent). Most of the time, 4 or 5 of those roads look to be dead by the end of the first turn - i.e. they lead to no intersection worth building on. The reason is that they place their starting roads toward the best available expansion position, but those positions get taken by other opponents during initial settlement placement. Oftentimes BOTH of the 2 available expansion intersections are taken, meaning that the AI initial roads literally lead to no location where an expansion is possible. Much of the time I would say that only 1 or 2 of the initial AI starting roads points anywhere potentially useable.

    This is apparently why a 4-player game (three AI opponents) is easier than a 3-player game (two AI opponents): more of their starting roads are unusable.

    When the AI opponents have poor expansion prospects (either none, or nowhere really worth pursuing), they buy development cards almost all the time. That in itself is not a poor choice for them to pursue. Combined with their tendency to ignore longest road bonuses however - as I mentioned in a previous post - the route to victory is typically quite clear. Build your 2 starting settlements reasonably close together, knowing that the AI opponents will ignore longest road. They will not pursue it themselves or interfere with opponents pursuing it. The overwhelming majority of AI starting roads will lead nowhere. In response they will buy development cards and typically obtain largest army among them. Meanwhile the human player places their original roads wisely so there are reasonable expansion prospects. Expand to 4 locations, build longest road along the way, and victory versus AI opponents will almost always come. The root problem is that conjunctively they have poor expansion prospects due to starting road mistakes. Oftentimes 2 of the 3 AI opponents have no expansion prospects at all. This problem combines with their tendency to ignore longest road to make victory rather reliable and easy.

    I have come to think that the major core weakness of the AI is the starting road problem, for reasons described. The best way to improve their overall performance in my view is to work on the starting road decision procedure. It plays itself out to become their core problem and weakness over the long term.

    A possible alternative: have the AI always point their roads towards the best available port. This procure would obviously be quite simplistic and still result in a considerable number of pointless starting roads, but I think they would still do better despite the primitive algorithm. Far more of their roads would lead to somewhere useful in a 4-player game then is the case now. Further possible improvements in their starting decisions would likely come to light with the primitive change suggested. One way or another I am convinced that the best strategic improvement overall for the AI will come from examining and working on their decision procedures for places initial roads. It is their biggest overall cumulative problem when all is said and done.

    I should emphasize that I do not mean to complain about the AI performance. AI scripts are extremely difficult to write. The simple fact that the AI procedures do not crash (result in a position where they make no decision) is hard in itself to accomplish. Their is nothing simple about it in the least. I am just looking to report the best observations about their weaknesses that I can.


  • administrators

    @NikkiR

    Thank you very much for your thoughtful and extensive feedback – it’s highly appreciated!

    I’ve forwarded this to the dev team as well as the QA-colleagues who are currently testing a new AI update as well. We agree that the AI going for the longest road is one of the areas where it needs improvement. Altering the choice of starting positions is another option. However tis will have to be discussed with the game’s inventor, with who we are in ongoing contact about the AI behaviour, as well.


Log in to reply