Decision tree
 

I have been playing around a little with decision trees. I entered MUD data into a decision tree maker algorithm. I used examples from 43 MUD entries here at TopMUDSites, and got the following tree from it:

The attributes used were codebase, worldsize, originality, pking, roleplay, theme. The question for the tree to answer is if the MUD is likely to have more than 100 players(true).

For example it states that if the worldSize is gigantic, and the codebase is rom where rp is accepted then from the data it concludes that the MUD has more than 100 players.

In some cases it has detected conflicting examples. They have the same attributes but not the came conclusion. To resolve that issue it use statistics.

The tree also shows some interesting patterns. For example that it is more unlikely the game has >100 players if there's unrestricted pking, but also mandatory rp seem to indicate less players.

What does everyone think? Does the tree portray the successful MUDs properly? Is there any interesting pattern in it, and is it true or just random luck?

Re: Decision tree
 

Oooers,
I love decision trees... and leafytrees in general (but thats beside the point) we use them at work sometimes.

I have a few questions about your methodology, because i of course dont know exactly how you generated your tree.. so ill just list some of the weaknesses of trees, and perhaps you have already counterbalanced them!

Because you are only inputting data to which you know the answer, you are making a "Training Tree" ... this tree can be used to determine the likliest category that an unknown data entry will have.
[That I mean by this is that you can use your tree to guess what population size may soon inhabit a mud that was newly created or one which you did not enter into the initial tree creation. ]

The unfortunate side effect to drawing conclusions from trees based on data such as the ones you inputted, is that you can "overfit" the data to the tree. You HAVE to get results similar to what you would expect, because you already know all the answers. (BTW, ime trying to paraphrase a book on this subject I read at work last week, if it seems garbled, ill retranslate when i go back to work after christmas ) There are ways to get around overfitting the data, but these do not take into account the conditions that may give rise to the properties you have inputted. For example, a VERY well written and interactive small mud may get a LOT of viewers, whereas a HUGE but miserably effected mud would soon vanish (and one must note that vanished muds are likely not included in your tree calcs... they may change things)

Also, it doesnt take into consideration advertising and website layouts... key things that I believe are quite important in attracting and keeping new players.... there was recently a post about someone's attempt to lure players with graphic intensive sites / client and keeping them after it was realized that the game was mostly text... That issue is also very interesting .

But beside that, I think its a really fascinating piece of work .. one definately worth doing, because it may give motivation to those littler muds to grow a bit more, and suggestions on Roleplay or PK status to others that may help their playerbase!

Re: Decision tree
 

I realized yesterday that the tree does show that it is somewhat overfit. To some unseen data it wont generalize well at all and simply not have an answer. I am considering methods to prune the tree to increase how well it generalizes.

Perhaps I should visit defunct list entries at e.g Mud Connector and MudMagic and add those as examples?

The problem is how to get such data. I guess I could include data if the MUD has an icon in the zMUD/cMUD client, but that would only cover a few MUDs.

Re: Decision tree
 

I read somewhere that a research project will almost always fail if you have a theory of the outcome in advance.

The reason being that the researcher would (subconsciously) only notice (and include) the parameters that correspond with his theory, while neglecting most of the ones that don't.

Re: Decision tree
 

I can't really say I agree with this. I work in a research field myself, and I can say from experience that almost all research projects undertaken in science are done with some hypothesis of what is likely to occur prior to the research being undertaken. This is why there are such exacting standards on how data is supposed to be collected, processed, etc. in order to avoid any implication of bias. So long as the results are repeatable and can be demonstrated as such, there's nothing to say the project has failed in any way, whether or not the final conclusions match the initial hypothesis.

Re: Decision tree
 

I agree with Galleus... All research projects have to have some sort of hypothesis to start with... and a lot of times its proven inconclusive by the evidence, or proven correct, However, this is why we have a peer reviewal process to ferret out those things that may make a project biased. It happens more often than one would expect that a paper is retracted or resubmitted to Science or NAture journals etc where a critical flaw has been revealed by another researcher that changes the findings.

Getting defunct listings from mudconnector would be a great way to increase the accuracy of the data, its important to get those failed stats in there too.