Warzone 2100 Forums

Hey all, I came across an interesting video regarding bot development in video games and thought it would be interesting to share since people here can relate to it. (Also because Siraj is a super good video maker). I’m studying machine learning and AI development IRL so at the moment I’m living and breathing this stuff all day. I wanted to tell you guys about what’s coming down the pipe for future AI development in general (not just video games).

https://www.youtube.com/watch?v=mGYU5t8MO7s

To summarize the video, there is new AI and machine learning technology available known as “deep” learning that has been around for a few decades but has only been implemented successfully as of 2012 (Geoffrey Hinton) and it’s a real game changer for the future of society. Basically instead of telling an AI or bot to do X, Y, Z based on some sort of conditions, we only tell it what we want (what winning looks like) and then by playing each game it learns from its mistakes and eventually converges to super-human levels.

The reason this learning is termed “deep” is because of how granular the information being analyzed is.

For example, instead of analyzing how many droids our opponent has, we can analyze every single pixel on the entire map which will not only give us how many droids our opponent has but will also tell us everything we need to know.

The way this would be implemented in Warzone would be to analyze every pixel of every frame. Warzone uses ~100MB of memory at any given time so an hour long game @ 30 FPS would be:
1 hour = 3600 seconds

(100x30) X 3600 = 10,800,000 MB’s (10.29 TB’s) which isn’t too bad.

Current convolutional neural networks with dual GTX 1080’s would train on a data set of this size for about 1 month to achieve good performance.

Anyways, just wanted to introduce deep learning here. Tensorflow is the current #1 machine learning framework right now.

It's a bit more complicated than that. But if you want to look at where the state of the art is today, just read this: https://www.theverge.com/2016/11/4/1351 ... e-blizzard

Per wrote:It's a bit more complicated than that. But if you want to look at where the state of the art is today, just read this: https://www.theverge.com/2016/11/4/1351 ... e-blizzard

Oh snap, nice. It's so ironic that you post that article because I've been contemplating how one would go about training a neural net on a RTS game, especially Starcraft because you can win money. I was curious if anyone was trying to pursue that so when I first saw the article my first thought was; already? or in the works? That article was published 7 months ago so I'm guessing in about another 6 - 12 months they'll present a Starcraft AI using a deep neural net. It's good to know that it's on Google's radar. Based on that article, they're essentially going to record tons of Starcraft games from the perspective of human users to tell an AI what winning looks like and what pixels are desirable (including mouse clicks). What I was suggesting is using all of the memory which essentially would be cheating... that's why I wasn't sure how one would go about training a neural net on an environment that's not entirely within view. Well, I suppose we'll see what Google can do since they are the world's leading organization in AI. We need to implement deep learning for Nexus. Call it Deep Nexus

MIH-XTC wrote:We need to implement deep learning for Nexus. Call it Deep Nexus

Hold on there, what if it becomes sentient and starts stealing my units and technology?

That's a very faraway future. However smaller projects, like microing units correctly in combat with the help of neural networks, may be much more reasonable nowadays, eg. https://www.youtube.com/watch?v=kW2q15MNFug