Jump to content
Sign in to follow this  
Kitten

Derp's thread of potential Machine Learning applications

Recommended Posts

Seeing as I keep putting this off later and later, I'm going to actually make a post to make myself come back and write up some detail about what this is and would do.

 

In basic terms, I would like to create a system that uses several algorithms in order to classify players by that latent variable, "skill". What this would mean is that even if one of the ratings would be paddable, presumably the rest of these algorithms would not be able to be exploited in the same way. Thus, the only way to get a truly high rating would be to actually be a good player.

 

Machine Learning (ML) is a fickle field, and it requires massive amounts of data for the alg. to even be considered viable. Since I will be doing this off of my own experiences, at this time I only plan on looking in to un/supervised learning ML algorithms. Just so you guys have some comparison, my last research in this area utilized 650,000+ annotated datapoints. I realize that having multiple algorithms working to ascertain a person's "skill" is much less effective in terms of serverload than the current WNx implementation (WN8) is, and I understand that. I merely plan to present my work as an alternative, perhaps it will be better in some ways and worse in others, I don't know at this time. If I identify an algorithm that is particularly effective at classifying players, it can undoubtedly be used much in the same way as WN8 currently is. That being said, it would be a part of a rating, not the whole thing, and somewhat defeat the original purpose.

 

As I mentioned earlier, I will need thousands of datapoints to make this work. Perhaps not quite as many as in my actual studies, but a fair amount. For some of the ideas in mind, I will need some community participation -- they will need to manually classify players and I will use it as a training set. I will need a core group within that set, the "experts", if you will, most probably superunica/unica to classify a couple dozen player's accounts. This is the "gold standard"; it is a form of quality control for the training set, which will be created by the community at large.

 

I realize that this post is already getting somewhat long, so I'll just leave it at "I'll give more details later". Hope this works out!

Share this post


Link to post
Share on other sites

The biggest problem you will face is getting not the number of datapoints but the number of fields you need. The API gives us a handful of useful fields such as battles, victories, caps, defs, frags, damage, survival, etc and we've all tried to ML new models out of those that do better than WN8; any success has been marginal. To make a breakthrough will require new data sources, I think.

Share this post


Link to post
Share on other sites

We'll see what happens Gryphon, the reason I am considering this is that the fundamental methodology will be very different from how WN8 was created. I'd be ecstatic if WG released more data; I will work with what we have though. Data will probably be a problem in terms of training, because for at least a few algorithms I want to tease out a rating (player specified) and then train the alg to classify any player inputted. There are some other ideas as well.

Share this post


Link to post
Share on other sites

I'm alive! As a first step for this, I think I will attempt to train/create an algorithm that determines the "true win rate" of a player (ie solo pub). I know RichardNixon has something of his own to discern expected WR, will be interesting to see how this compares. At the moment I can only think of how to use account/stats for this rather than tanks/stats. Sorry guise! I'll try to figure out a way to make that work. In the meanwhile, I should have more time soon, so I hope to get this done in such a way that it can be implemented server-side for websites like wotmanager, if there is interest. I'd assume recruiters would be happy to have a tool like this.

tl;dr I'm working on getting a system that would discern a player's true win rate from their overall account stats. I want to be able to get it for individual tanks too... In the future.

I will e-beg for some volunteers to judge people's wr soon as well.

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Recently Browsing   0 members

    No registered users viewing this page.

×
×
  • Create New...