Nov 302015

While cruising through talks on youtube I found one titled “I am a legend: Hacking Hearthstone with machine learning“, an interesting application of machine learning in video games:



Hearthstone is a card game in which each player has a deck of characters, the goal of which is to use those cards to do damage against your opponents. I digress, the essence his talk had two major themes: finding imbalances in cards predicting an opponents moves. The former involved deriving an over determined system and simply applying least squares, the ladder involving machine learning. I think a simple explanation of what his algorithm does is extract sequences of cards played from a rich data set:

Experimenting with various dataset sizes, I started to get consistent results when using over 10,000 game replays. Yet using more than 50,000 games does not improve the results, so I settled for a dataset of 50,000 replays that were played between May and July 2014.

Part of writing a good talk is being able to balance technical and non technical aspects of a subject while maintaining the attention of audiences whose background is varied. I think Elie did a decent job of this, though of course I would be bias and hope for a bit more depth in the machine learning.

I was a bit upset that the speakers said they were going to release their software and then didn’t, but their reasons were sensible after contacting the creators of Hearthstone. However there are parts of their software that I think they should release, such as the front end web application and back end webserver, this would allow people to write their own statistical/machine learning plugin. Providing the machine learning portion of the software is what gives users an unfair advantage, which is what the creators of Hearthstone were concerned about.

Another extremely interesting application of machine learning to video games is MarI/O , an application of creating a neural network which can pass levels in Mario:



The author of MarI/O, youtube personality SethBling (I’m not sure what his real name is), used an application of neural networks. Essentially the neural network is a mapping of the buttons you can press to move your character to blocks on the screen and then associating a distance travelled based on that mapping. If Mario makes it further in the level then a previous mapping, then his fitness goes up and he is evolving. As an aside, neural networks are the same technology that companies like Nvidia use to identify vehicles. Jonathen Cohen, former director of the deep learning department at Nvidia, describes the methodology in a short demo.

These are only some relatively recent talks on machine learning in video games, there are certainly other applications and more we’ll see in the future.

 Leave a Reply

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code class="" title="" data-url=""> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong> <pre class="" title="" data-url=""> <span class="" title="" data-url="">