how game developers are using stats + data mining

[joeclark77]

My academic work is in the area of business intelligence and big data.  I was browsing articles about the R language (at R-Bloggers.com) and came on this interesting link.  It's a presentation about the use of analytics the gaming industry called "Knowing How People are Playing Your Game Gives You the Winning Hand".

Obviously I don't have a game sophisticated enough for this, but there's some interesting ideas in it.  For example, you might try to build a model of user behavior to adapt the AI.  Or you might analyze where people have difficulties, or what interventions might prevent somebody from dropping the game.

Worth a look?  Get any inspiration?

http://blog.revolutionanalytics.com/2013/06/how-big-data-and-statistical-modeling-are-changing-video-games.html

[Eben]

Noooooo, not R!!!!

I've recently been touch in inappropriate ways by R

The actual collection of in-game data and player modeling is something I've been working on myself. There's lots of really good ways it can be used, such as shifting loot drops in favor of how the player prefers to be playing.

Also, don't use R.

[guest509]

I prefer Brainf**k

http://en.wikipedia.org/wiki/Brainfuck

EDit: I apologize, i have nothing to say on the original topic. I just like the idea of that language is all...

[Etinarg]

"Knowing How People are Playing Your Game Gives You the Winning Hand".

While the idea is right, I don't think data mining is a good approach to solve the problem. In my opinion the best idea is to get someone new to your game, ask them to play, and watch them playing. This way you'll see where they have problems, where they have complaints, get feedback about what they like and of course you can see the strategies they try and use.

I'd say it's hard to replace this by data mining, becuase watching the player gives way more information, also on an emotional and empathic level, than checking logs and saved data.

[Quendus]

Just do both. Or whichever you have the resources to implement. Some people don't have web servers that can receive the data, some people don't write network code, some don't have RL friends that are willing to play a roguelike without a gun to their head.

[Eben]

It may not be entirely clear, but the data mining is typically used after player models are understood already.

So the steps are:
1) Model the types of players (keeping in mind your preferred type for making money)
2) Gather data on everyone playing
3) Mine the data to see what matches the various player types
4) Make adjustments suggested by the data mining to turn players into the ideal type
5) Repeat until profitable or broke

For a small user base it's not feasible to model actual player types, but if you have an idea of how your game should be played (hint: you should) you can see if it is being played that way and maybe adjust. Just don't expect huge results without a huge set of data.

[joeclark77]

I think data mining might be particularly valuable with MMOs where you want to understand the emergent behavior that you can't get from interviewing individual players.  EVE-Online does a lot of analytics that they share with the community, and I know they use them to try and figure out which ships are overpowered, how asteroid minerals might need to be balanced, what's going on with the in-game economy and so forth.

[joeclark77]

Noooooo, not R!!!!

I've recently been touch in inappropriate ways by R

...

Also, don't use R.

I actually offered a college course on data visualization with R this spring.  (Didn't get enough signups to go ahead with it though, so the minds of the young were spared this time.)

It's not for everybody, but it's got some very neat features and lots of packages are available if you're into data analytics and visualizations.

I've used it to program computer simulations for research papers, and it worked well enough to get my dissertation committee to sign off on my doctorate, but it seems to eat memory or slow down for some arbitrary reason when you do a lot of repetitions, so it's not the best tool and I'm looking into Python now.

[Krice]

I think the result of data mining can be an average of everything. Sucks.

[Quendus]

Fit a multi-modal distribution, if you think that a single mode can't describe the data accurately.

[Eben]

I think data mining might be particularly valuable with MMOs where you want to understand the emergent behavior that you can't get from interviewing individual players.  EVE-Online does a lot of analytics that they share with the community, and I know they use them to try and figure out which ships are overpowered, how asteroid minerals might need to be balanced, what's going on with the in-game economy and so forth.

If there's one thing I've learned from watching Let's Play videos it's that the things you hear people talk about in games when asked later (or self volunteered information) is not at all the same as what they run into while playing. There's so many little things that happen that they either forget or just don't think is a big deal.

OTOH, if you've got an achievement system you're already data mining, and if you're going to data mine you should add achievements