While this is somewhat off-topic, I figured some of you might be curious enough to help me out.
I am doing a research project (master thesis) where I attempt to use machine learning and game play data collected from a modified open-source version of Super Mario. The hypothesis is that men and women play differently and that this difference can somehow be extracted as features used to learn a "gender classifier".
One can imagine that it may be possible to transfer the concept to other software applications and to extract other information than gender about users.
I am curious what you think about the project, and if you have time: play the game for 5 minutes and generate some data.
The game client is Java-based and can be found here:
http://www.masterofdebian.com/mario
/Lasse
Gender prediction based on information about how a person plays a game
I tried three times, but the controls seemed to work sporadically: sometimes having no effect, sometimes seeming to get 'stuck down', sometimes taking an action I didn't press the button for (or, perhaps, doing what I asked but with a several-second delay). Hopefully this didn't generate too much spurious data for you, sorry!
Author
I tried three times, but the controls seemed to work sporadically: sometimes having no effect, sometimes seeming to get 'stuck down', sometimes taking an action I didn't press the button for (or, perhaps, doing what I asked but with a several-second delay). Hopefully this didn't generate too much spurious data for you, sorry!
That is perfectly fine! I agree that the game controls are not superb, and I did not program that part myself. The game is a modified version of the MarioAI from http://www.marioai.org/.
There should be no delay generated by submitting game play data as it occurs on a separate thread.
This is a very interesting subject study, I would love to see the results when it is completed.
I think the trick will be controlling for other factors. For example, the male subjects may be a higher proportion gamers than the female subjects, which may skew results.
Interesting topic. Please post a copy of your paper once it's done. I have done similar work analyzing blog postings and newspaper articles.
My only advice would be to think about including context in your measurements. Rather than just controller inputs, consider measuring percentage of bricks smashed, percentage of time spent on ground, number of times player got within n pixels of an enemy without killing the enemy, percentage of enemies squished, etc..
Also, breaking each level up in to distinct areas, and keeping statistics about each area, might also help distinguish male from female players.
My only advice would be to think about including context in your measurements. Rather than just controller inputs, consider measuring percentage of bricks smashed, percentage of time spent on ground, number of times player got within n pixels of an enemy without killing the enemy, percentage of enemies squished, etc..
Also, breaking each level up in to distinct areas, and keeping statistics about each area, might also help distinguish male from female players.
I am doing a research project (master thesis) where I attempt to use machine learning and game play data collected from a modified open-source version of Super Mario. The hypothesis is that men and women play differently and that this difference can somehow be extracted as features used to learn a "gender classifier".[/quote]
Immediately, my question would be "to what end"? As long as this information isn't empirically true (which it never is, when dealing with people), it's pretty useless. There's already a vast understanding about how women play games compared to men, but it's highly generalizing and you cannot use it to figure out the gender of the person.
To put it simply: Too many men play like women and too many women play like men.
As for how women actually tend to play compared to men, it's mostly about attention span. Women tend to have a playstyle that reflects how they appear to the outside world (such as being diplomats, managers and strategists), whereas men tend to play in terms of how the world look to them (down-and-dirty, completionists, tacticians, bragging rights, etc). But that's pretty much it. The rest is pure individuality.
Anyways, just my 2 cents.
Author
Interesting topic. Please post a copy of your paper once it's done. I have done similar work analyzing blog postings and newspaper articles. My only advice would be to think about including context in your measurements. Rather than just controller inputs, consider measuring percentage of bricks smashed, percentage of time spent on ground, number of times player got within n pixels of an enemy without killing the enemy, percentage of enemies squished, etc.. Also, breaking each level up in to distinct areas, and keeping statistics about each area, might also help distinguish male from female players.
I will post my paper here when it is done later this summer, and thanks for the contributions so far.
I do collect the variables you mention and many more, except how close players get to enemies without killing them; it is a feature I will try to add, since I can use the other data gathered to produce that particular feature.
Author
To what end? To finish my master degree obviously 
But my focus is rather on applying machine learning techniques to solve the problem. First and foremost I do not have any ideas as to what features will help classify gender the best other than guesses, and I will leave that to the computer to learn on its own. IF classification is possible, the next step is to apply unsupervised learning to the problem, which is far more interesting (if working) than supervised learning.
But my focus is rather on applying machine learning techniques to solve the problem. First and foremost I do not have any ideas as to what features will help classify gender the best other than guesses, and I will leave that to the computer to learn on its own. IF classification is possible, the next step is to apply unsupervised learning to the problem, which is far more interesting (if working) than supervised learning.
I am doing a research project (master thesis) where I attempt to use machine learning and game play data collected from a modified open-source version of Super Mario. The hypothesis is that men and women play differently and that this difference can somehow be extracted as features used to learn a "gender classifier".
Immediately, my question would be "to what end"? As long as this information isn't empirically true (which it never is, when dealing with people), it's pretty useless. There's already a vast understanding about how women play games compared to men, but it's highly generalizing and you cannot use it to figure out the gender of the person.
[/quote]
You can't use your imagination a little? There are all kinds of reasons you might want to know the gender of the person sitting in front of a machine. Targeted advertising, for one. Maybe targeted rewards (a pink shotgun instead of a camo one?), or even changes to storylines based on how a player is percieved to act.
And there's always the "plays like a girl" badge on the PS3 that's currently not an option.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
