Single or Multiple Behavior Trees?

Question

I just finished coding a generic Behavior Tree structure for my games. My question is, when creating behaviors for enemy AI's, do I create one large behavior tree with every possible configuration as a node or do I create many multiple Behavior Trees and simply swap them in and out as I need them?

To better understand my question here are some examples:

One large Behavior Tree might look something like this:

Multiple smaller Behavior Trees might look like this:

In the second tree I would simply swap out different trees depending on when I need them. This is more apparent for things like evading and backing away from obstacles, I feel like that would need to be done alongside many other behaviors like chasing and evading.

I hope that made sense.

Answer 1

Generally there would be no need to "swap" out trees.

Build all your behaviors, add in your sequences and selectors to make your tree for your ai agent.

Then for special case stuff you add decorators to stuff where needed.

One big tree, per agent "type" so a soldier type would have a tree with patrol, attack etc etc and a miner would have a different tree. They would use many of the same objects (ie. go to way point) but you just wire up your tree for each agent you needed.

The first picture you have in your post is the way to go.

Answer 2

Not sure if I understand your question correctly. Do you mean during runtime? Or when you design your trees? Certainly when designing your trees, it helps a lot to create lots of small, reusable, modular trees.

You can try creating the concept of "tree link" nodes. Nodes that just let the traversal jump to a different tree (a whole separate tree, stored in a different file, perhaps), and when that tree finally gives a return value, uses that as its return value.

This way you can have a generic main tree that have branches touching on broad subjects: survival, attack, idle, etc. Then you just add your tree links into those empty branches to create variations. You may have an attack subtree that's designed to keep away and just do ranged attacks. You might have another one that's designed to charge head on and do melee attacks instead.

This way its easy to organize your AI types, by just doing a mix-and-match approach. This is not necessarily for convenience in modifying your trees during runtime, but that can be possible too.

Doing this also makes your trees a lot more readable to fellow teammates. It's essentially using the concept of "abstraction" and "information hiding". Two very powerful concepts in software engineering in helping to keep large systems easy to understand and maintain.