To program a AI for this game is hard. So many choices to make and then keeping in mind the next turn, include the random dice factor?
Making a AI that interprets a (RL)humans next move and dice randomness?
Totally untrue. There exists proficient AI for games much more complex than Blood Bowl. The idea that complex games require complex AI is mostly a fallacy that comes from non-programmers and which is based on the idea that game AI has to "think" like a human does - it really doesn't.
I think one of the AI mod files did all that, and I know a list of suggestions for player "type" changes was sent to Cyanide in the past, but they didn't do anything with it.
Actually, they incorporated the XML changes into LE, according to discussions from way-back-when.
But isn't it possible to modify the AI behaviour using these parameters? Even if it is only general behaviour. I know it won't be possible to create an awesome AI, but maybe we could try to decrease the amount of GFI the AI does, for example. Or limit dodging to AGI4 teams. I don't know, something like that.
Reducing the risk that the AI takes just makes for boring games in which the AI never scores. The AI already never wins, but it will occasionally score a point due to taking big risks. There's nowhere to take the XML variables that they haven't already been taken years ago, though you can no doubt get a lot of play by randomly changing numbers and letting the placebo effect take you places with your testers. As of LE, many of the XML files have no effect on the game, having had their values hardcoded during development but never returned to pulling from the data files.