I-KP, thanks very much for the excellent reply. I'm not offended in the slightest, no worries there. And, I think that you did make your point much clearer to me; I confess I had missed the points about being sometimes-on, etc, in your other notes. I am quite verbose, but I will try to keep this post brief; it's out of time consideration, not out of any sort of sense or irritation or whatever.
You are also spot on in the fact that this is not a learning AI in any fashion, and that high-order tactics are essentially random occurrences. Hence the nature of many first-and-second-order emergent systems, but it will never evolve beyond that. Learning AIs are a whole other school of thought which is beyond my threshold of interest at the moment.
I do take issue with the assertion that ships always act purely selfishly in AI War, though, as they do give some weight to what other ships are doing. That leads to groups spontaneously forming or dissolving, rather than just ships purely winding up together by chance.
The other founding tenet of the AI in AI War is that it really doesn't matter what the AI is doing, it just matters what it makes the human players do. From that angle, I am wholly disinterested in improving the AI for the sake of improving the AI, but only in terms of making new and interesting decisions and challenges for the human players. Having the AI appear more human-like by using combined arms is not too much of a goal. Having challenging, varied new formations by having it use combined arms is a worthy goal.
Part of this, then, comes back to what is computationally possible in a game of sufficient complexity. Or, rather, what can be algorithmically designed in a linear fashion by programmers, without introducing other seemingly-unrelated problems. Discussion of combined arms is all well in good, but at some point that has to actually be modeled in a mathematical sense, which is far easier said than done. In having thought about this issue in more depth over the last couple of days, I think that I am coming closer to a model for making this work.
The problem is, when designing an emergent AI it really can't be based at all on what we as meat-sacks would do. The more we try to emphasize meat-sack-style rules, the more the effectiveness of the underlying AI crumbles in general. Just now, I have been struck by the thought that it might be good to try enhancing the flockability of the AI in general, which would lead to more self-organizing combined arms of all sorts. Forget starships, it could work on an interesting level with many ships; though anything with munitions or similar boosters could certainly be a higher priority for this sort of thing. The common way that RTS AI tends to solve this is via an "anchor unit," but there are a ton of pitfalls with that on anything smaller and less sturdy than a starship. Better, potentially, to organize around mutually-selected targets, with starships and other boosters then perhaps acting as tag-alongs to targets with a lot of other ships assigned to them. I will have to play with that some, and see what I can come up with.
P.S. -- As to why an AI starship is ever best alone, the reason is simple: distraction. A starship alone is able to take out a lot of defenses and things players want to protect while on its own, and so the tendency is for players to then run after that. On the flip side, they take a long time to kill, so that can then let all the other ships slip past and do a bunch of damage elsewhere in the meantime. In the end, that can lead to greater losses for the players in the end. Not a good tactic to use if you are worried about losses you take, but in terms of creating interesting and challenging situations for the human players, I think it succeeds pretty well.
