What is the best way to implement this &#34;weird&#34; kind of graph in C++ for a game?

Michael Moldenhauer · 2012-02-27T23:53:20

Hi. I've asked this on a couple of other forums and got some useful responses but I'm still not 100% sure as to the best way to proceed with this. I've got this problem of representing a "weird" kind of graph structure, which is like a normal graph but it has an additional layer of structure called "strings" which are composed of nodes connected together and ordered in a linear order. The purpose of this graph is to represent a system of levels in a game. In one of the responses I got, it was suggested I should implement this using a conventional graph type (i.e. just an implementation of a conventional graph, like Boost's graphs). But the problems appear when considering the "strings". Namely, I have something like this: class WeirdGraph { private: class NodeString { private: <...> std::vector<WeirdGraphNodeIterator> nodes; <...> public: <...> }; <... (there is a graph variable declared here) ...> std::vector<NodeString *> strings; <...> public: <...> }; i.e. the weird graph object contains a private class to represent the strings. Problems are that: 1. iterators in the graph get invalidated. Each node string stores the nodes comprising it as iterators into the graph. When a node is deleted, the iterators get invalidated, and this seems to require an update to the NodeStrings. What can be done? Obviously, one could just use a graph that does not invalidate iterators (though this may lead to a penalty accessing/traversing the graph, hence why I'm a little hesitant, since I'm going to be accessing more than I'm going to be deleting stuff, though I doubt this is used in performance-critical code in any case so it probably doesn't matter.). But I'm curious to know if there may be a technique that would work even in the case of invalidating iterators. 2. To update, the node strings need to know what nodes belong to them. So the nodes contain a string ID field, which is user-specified when the user creates a string and the WeirdGraph enforces uniqueness. Is it "good form in C++" to have WeirdGraph do the enforcing and not NodeString? It feels iffy. 3. Regarding that, is there an alternative to using the user-specified string IDs that won't fail (e.g. incrementing a counter may fail long off into the future, but I doubt one will be creating and deleting billions of times. But is having that "latent failure" possibility in there OK? As it feels iffy too.)? 4. NodeString constructs nodes in its constructor, then deletes them in the destructor. This last bit makes me feel iffy as then it's not okay to pass one by value -- though we can always not do that, but I'm not sure whether having such requirements which are enforced only by user understanding is a good idea (although does it matter here since NodeString is a private internal of WeirdGraph?). Any comments on these issues? Right now I'm really leaning toward using a non-invalidating graph type, as that would entirely eliminate 3 out of the 4 issues above, but I'd still like to know what the deal is on the 4th. The 4th is also why I'm using pointers above (perhaps I should use std::auto_ptr instead of raw pointers there? hmm) instead of objects directly, since that would necessitate some copying/by-value passing when filling that vector.

swiftcoder

18,997

February 11, 2012 12:19 AM

Would the best approach, then, be just to use some kind of graph that does not invalidate iterators on deletion, or has some other type of stable node reference?

You just told me that nodes don't ever get deleted, only strings get deleted. In which case, node iterators can never be invalidated. Which is it?

Tristam MacDonald. Ex-BigTech Software Engineer. Future farmer. [https://trist.am]

mike3

178

Author

February 11, 2012 02:26 AM

[quote name='mike3' timestamp='1328918404' post='4911825']
Would the best approach, then, be just to use some kind of graph that does not invalidate iterators on deletion, or has some other type of stable node reference?

You just told me that nodes don't ever get deleted, only strings get deleted. In which case, node iterators can never be invalidated. Which is it?
[/quote]

Deleting a string deletes the associated nodes. That happens in the string object's destructor. If you were thinking of this post:

http://www.gamedev.net/topic/619851-what-is-the-best-way-to-implement-this-weird-kind-of-graph-in-c-for-a-game/page__view__findpost__p__4910656

I was trying to say you could not delete nodes individually, not that they couldn't be deleted at all, as deleting a string entails deleting all the nodes that belong to it. Sorry if that wasn't clear.

swiftcoder

18,997

February 11, 2012 02:43 AM

Deleting a string deletes the associated nodes. That happens in the string object's destructor.

But surely, the same node can occur in more than one string? If so, what happens to the other strings when a node they depend on is deleted by another string?

If not, then your strings are disjoint subsets of the graph, and there is no particular advantage to storing the nodes in a shared container.

Tristam MacDonald. Ex-BigTech Software Engineer. Future farmer. [https://trist.am]

mike3

178

Author

February 11, 2012 04:41 AM

[quote name='mike3' timestamp='1328927187' post='4911857']
Deleting a string deletes the associated nodes. That happens in the string object's destructor.

But surely, the same node can occur in more than one string? If so, what happens to the other strings when a node they depend on is deleted by another string?

If not, then your strings are disjoint subsets of the graph, and there is no particular advantage to storing the nodes in a shared container.
[/quote]

Yes, they are disjoint. But they can be connected together by connecting nodes from different strings with an edge, and thus I use one graph, to represent the whole "web" of nodes. And also, because building on a standard graph type was suggested on one of the other forums I asked this question on, as it's a "standard" kind of data structure.

Antheus

2,410

February 11, 2012 03:22 PM

::iterator types are for iteration. While some may be held on to, it causes aforementioned problems.

In most newer languages, iterators are explicitly checked for modification and will cause failures when trying to use them as references.

Iterators are not identity types and cannot be used for such purpose, even though in some corner cases, depending on implementation, they can temporarily take such role.

Instead of insisting on such approach, use a different approach. Establish strong identity (NodeID, StringID, EdgeID). Keep track of those separately as well as their connections. Then build structures on top of that.

There really are no difficulties here. It's just a basic list/tree structure, which iterators make impossible to implement if any of them can be mutated, something that iterators, as a concept, cannot handle. Only general solution to this problem is garbage collection, which doesn't solve cross-linking issues.

mike3

178

Author

February 11, 2012 09:07 PM

::iterator types are for iteration. While some may be held on to, it causes aforementioned problems.

In most newer languages, iterators are explicitly checked for modification and will cause failures when trying to use them as references.

Iterators are not identity types and cannot be used for such purpose, even though in some corner cases, depending on implementation, they can temporarily take such role.

Instead of insisting on such approach, use a different approach. Establish strong identity (NodeID, StringID, EdgeID). Keep track of those separately as well as their connections. Then build structures on top of that.

There really are no difficulties here. It's just a basic list/tree structure, which iterators make impossible to implement if any of them can be mutated, something that iterators, as a concept, cannot handle. Only general solution to this problem is garbage collection, which doesn't solve cross-linking issues.

However, I've noticed in some implementations like "Boost" that there are "descriptors" in addition to iterators, but at least in the case of Boost, with some forms of the graph these too can be invalidated, and so we get the same problems. Would I be right in guessing that these problems are saying that messing around with invalidating references is the wrong way to go, and there's simply no way to use invalidating references in this way without making code a mess? As it seems unavoidable, and so I'm really strongly thinking of using your ID-tag approach. Especially considering I was already going in that direction with giving strings ID codes by which one could reference and access them. The only downside I can see is the added time complexity of the accesses, but considering that this is not performance-critical code that is not actually a problem.

mike3

178

Author

February 20, 2012 10:26 PM

NEW: Well, now I've settled on the approach of using a graph that can refer to nodes by ID codes instead of iterators. This removes problems 1, 2, and 3 in my original post, but I'm still not sure what to do about problem 4. Is it okay for the "string" class to be "unsafe to copy" due to how it creates/deletes iterators in the graph (see the OP)?

LorenzoGatti

4,648

February 21, 2012 09:33 AM

::iterator types are for iteration. While some may be held on to, it causes aforementioned problems.

In most newer languages, iterators are explicitly checked for modification and will cause failures when trying to use them as references.

Iterators are not identity types and cannot be used for such purpose, even though in some corner cases, depending on implementation, they can temporarily take such role.

This is how iterators work in the real world: their lifespan is limited to some activity that iterates through members of a data structure, and in this small scope one can take care not to invalidate them with inappropriate operations.
If you insist on storing iterators instead of plain pointers or names, you are abusing the standard library and looking for trouble.

Omae Wa Mou Shindeiru

mike3

178

Author

February 27, 2012 11:53 PM

[quote name='Antheus' timestamp='1328973736' post='4911974']
::iterator types are for iteration. While some may be held on to, it causes aforementioned problems.

In most newer languages, iterators are explicitly checked for modification and will cause failures when trying to use them as references.

Iterators are not identity types and cannot be used for such purpose, even though in some corner cases, depending on implementation, they can temporarily take such role.

This is how iterators work in the real world: their lifespan is limited to some activity that iterates through members of a data structure, and in this small scope one can take care not to invalidate them with inappropriate operations.
If you insist on storing iterators instead of plain pointers or names, you are abusing the standard library and looking for trouble.
[/quote]

Actually, I had programmed my own graph implementation and featured iterators. So I guess I was abusing the concept of iterator, then. But there's no problem now -- I've since replaced that graph with a new version that uses ID codes to give stable references, like Antheus suggested. Granted, in theory the use of std::map gives a performance hit, but this code is not used in anything performance-critical so that is no problem.

What is the best way to implement this "weird" kind of graph in C++ for a game?

This topic is closed to new replies.

Popular Topics

Recommended Tutorials

What is the best way to implement this "weird" kind of graph in C++ for a game?

This topic is closed to new replies.

Popular Topics

Recommended Tutorials

Reticulating splines