?????

A faster way of doing this would probably be to have two different trees, one with weights of one type and another with weights of another type. This will ensure that your tree splits are more optimal, and you can use completely different dimensions as well without a slowdown.

Interesting idea! If I can spare the cycles, that would be interesting to do. Are there segmentations that are useful against adaptive movements but utterly superfluous against non-adaptive movements? (other than data age?) Vice versa?

I'm unsure if that would be faster really. Since these trees are generated on-the-fly without rebalancing, it's not like they're hugely optimal in the first place. I doubt the overhead of changing the weighting would be much compared to how non-optimal it is in the first place. Additionally two trees could be worse from the perspective of on-CPU caches. IMO it's worth trying both ways and benchmarking.

The end of the round probably would be a good time yes, though I've never really tried it. The re-balancing would take longer each round though, so it may make sense to only do that for the first few rounds of battle. As the number of rounds increases there would be diminishing returns to constantly re-balancing anyway.

The first few rounds makes sense.. Relatively quickly your dimension balancing should conform to the movement pattern of the enemy, and later rounds probably wouldn't shift the balance dramatically.

The distance dimension of a bot that heavily prefers a certain orbit distances will get properly split after just a few re-balancing operations, making further balancing on that dimension a comparative waste of time I would imagine.