Jump to content

  • Log In with Google      Sign In   
  • Create Account

We're offering banner ads on our site from just $5!

1. Details HERE. 2. GDNet+ Subscriptions HERE. 3. Ad upload HERE.


Don't forget to read Tuesday's email newsletter for your chance to win a free copy of Construct 2!


Behavior Trees


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
4 replies to this topic

#1 rpiller   Members   -  Reputation: 690

Like
0Likes
Like

Posted 23 May 2012 - 07:16 PM

I have a question specifically around giving actions access to the data it needs. Since tasks are their own classes, and my player, enemy, and map data are in their own classes, what is the generally accepted way of letting these tasks in a behavior tree have access to all this game data? With the idea of no globals anywhere (a rule I generally like to follow) how should they be getting access to this game data information?

Are we literally passing all this data to every node starting with the top node which will do the passing of all this data to it's child nodes as it transverses the tree? That doesn't seem extendable as new things need to be created and passed to each node. Are we creating another class that holds all these things and passing that? Seems extendable but sort of cheap to give each node access to things that that specific node doesn't/shouldn't care about. For example if a tasks job is to check if the actor has ammo before going onto another task to shoot the gun, that task doesn't need access to the player information or the map data. It would just need access to the actor that tree is attached to.

Sponsor:

#2 jefferytitan   Crossbones+   -  Reputation: 2222

Like
0Likes
Like

Posted 23 May 2012 - 09:45 PM

That's an interesting question that I haven't seen posed before. In general I would say that as behaviour trees are often aimed at a game designer (not a developer), it is dangerous to filter the available data too much, as the designer may be unable to add the kind of nodes they want further down the tree if the information required by those nodes is not available.

Having said that, global variables or passing huge state objects is not good either. Perhaps some kind of blackboard or factory object which can obtain a node the data that it needs? Or alternately you have node types, which only get a certain type of data, but that limits the capabilities somewhat.

#3 rpiller   Members   -  Reputation: 690

Like
1Likes
Like

Posted 24 May 2012 - 11:57 AM

Is there any generally accepted method for this?

I'm very new with Behavior Trees, but I was playing around with the idea of these tasks & actions being functions inside the Gameplay state class which would hold all your game data anyway. Each node would be an event which points to a function you create in your Gameplay state class instead of being their own classes in themselves. When learning about these I find it very odd that actions or tasks are their entirely own class because it seems really all it needs to be is a function. Then all the nodes in your tree just point to the function to run instead of creating an entire object. You'd still need a tree hierarchy setup but they could only store variables of their children, and basically a pointer to the function to run isntead of definding their own Update() methods. I'll try to get this in practice soon, but it just seemed odd that tasks and actions where their entire class. Seems by doing that you have to break/bend some "guidelines" with programming to get them to see the data they need.

I'm a big OOP guy so I can see breaking them into classes helps separate the creation of these easier then if they were all in the same class, but it still just seems odd the tasks/actions are entire classes themselves.

#4 LorenzoGatti   Crossbones+   -  Reputation: 2736

Like
1Likes
Like

Posted 25 May 2012 - 02:18 AM

The natural place for obtaining and maintaining the data needed by behaviour tree nodes (or other kinds of AI) is the game state representation. For example, queries like what entities satisfying some criteria are close to a certain map location are a job for the main spatial indexing system; they could be exposed to AI clients as a service offered in a general form by a game state façade, or splintered into specialized methods of specific entities (e.g. enemies that a unit can see or treasures that a unit can pick up in less than x ticks).

Since a behaviour tree is the behaviour of something, it will always execute in the context of the appropriate game entity, and it can make rather strong assumptions on that entity's type and state: all of that entity's query methods are at its disposal, and the rest of the game state should be easy to reach..
Produci, consuma, crepa

#5 jjd   Crossbones+   -  Reputation: 2117

Like
0Likes
Like

Posted 28 May 2012 - 10:42 AM

Is there any generally accepted method for this?

I'm very new with Behavior Trees, but I was playing around with the idea of these tasks & actions being functions inside the Gameplay state class which would hold all your game data anyway. Each node would be an event which points to a function you create in your Gameplay state class instead of being their own classes in themselves. When learning about these I find it very odd that actions or tasks are their entirely own class because it seems really all it needs to be is a function. Then all the nodes in your tree just point to the function to run instead of creating an entire object. You'd still need a tree hierarchy setup but they could only store variables of their children, and basically a pointer to the function to run isntead of definding their own Update() methods. I'll try to get this in practice soon, but it just seemed odd that tasks and actions where their entire class. Seems by doing that you have to break/bend some "guidelines" with programming to get them to see the data they need.

I'm a big OOP guy so I can see breaking them into classes helps separate the creation of these easier then if they were all in the same class, but it still just seems odd the tasks/actions are entire classes themselves.


Good question. Having behaviors that are simply functions is appealing in several ways. One of the main benefits is that you are obviously using a functional approach so there should be no side-effects, i.e. when presented with exactly the same data the behavioral should produce exactly the same result (unless of course you intentionally introduce randomness). However, classes provide greater flexibility. There is nothing to prevent you from using them in a functional way, but they also provide a way of retaining state if that is useful, .e.g if there is data that you want to make available to the behavior when it is created or maybe the behavior has a cool-down period. Class-based behaviors also provide a way to reuse code when the behavior is parameterized.

Having the behaviors use functions on a game state object is, IMHO, a reasonable way to go. However, I would suggest using a class, say, BehaviorContext, that provides the interface for the behaviors to use rather than the whole game state. This interface helps to focus on what information should be available for behaviors rather than exposing all information.

-Josh

--www.physicaluncertainty.com
--linkedin
--irc.freenode.net#gdnet





Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS