I was originally planning to do a post on brute-force methods, but that idea flew out the windows when I noticed this article in my weekly newsletter: GTA V + Universe. The article announces GTA V support for a library that you can use to train your AI in games.
What in the universe are you talking about?
Obviously, this is great. Just in case you got lost in how great this is, I’ll give you some context. Universe is a Python library for training agents using games. It interacts with the games in much the same way that a human would, as OpenAI explains on their website:
Universe allows an AI agent to use a computer like a human does: by looking at screen pixels and operating a virtual keyboard and mouse.
If you’re from my generation I’m sure you know Grand Theft Auto (GTA). It’s an open world sandbox game. The developers have created simulations of cities, with sometimes ridiculous depth and complexity. The cities are humungous and densely populated. There are millions of NPCs who all have quite intricate AIs, cars, traffic, pedestrians, shops, minigames and of course weapons. In GTA V the devs really out did themselves with the degree of realism. Usually, GTA games usually feature a lot of crime but that is far from relevant in this case.
Why is Universe big?
Machine learning is awesome (like I discussed already in my AI 101 post). With it, you can teach computers to do things better than you could ever hope to achieve yourself. The drawback to this is that you have to train them. This training process is usual a lot longer and more intensive than training any human. The agents are also much more likely to exhibit extreme and unpredictable behaviour during the training process. This means that training AIs can be a big bottleneck in research.
Say for example that you want to train a self-driving car. That means that you need, at the very least a car, interface equipment to let your PC talk to the car, multiple other cars to simulate traffic and a safe space to test it in. That can get very expensive. With Universe researchers potentially have a much safer and cheaper way to train and test their AIs. (You do need to buy a copy of the game you want to train you AI in, but 60$ is much more manageable in comparison to what I just mentioned.)
Obviously, this is great for training Game AIs. A lot of AI in games is usually not very good. I think it would be interesting to see if people can use this to come up with better AIs to play against. I’m much more excited, however, about the possible explosion of AI research. With safe, accessible and relatively inexpensive ways to train your AI, I hope that this will start a wave of research. The reason I’m excited about this project is the same reason I think that sexism and racism are bad: if more people are allowed to participate, you get better innovation and that benefits everybody.
Why should I care about GTA?
That is why it is great that Universe exists, but I also want to touch on the fact why it is awesome that they got to support GTA V. Like I said, GTA V is absolutely massive. The possibilities for the different kinds of AI’s you could train there is mindboggling. Like I said there are things like traffic to train your car AI, but there is so much more. Collision detection, object recognition, natural language processing, flight AIs, the list goes on. Another great fact about GTA is that it all comes within the same game. Like OpenAI mentions on their website:
However, despite all of these advances, the systems we’re building still fall into the category of “Narrow AI” — they can achieve super-human performance in a specific domain, but lack the ability to do anything sensible outside of it. […]If we are to make progress towards generally intelligent agents, we must allow them to experience a wide repertoire of tasks so they can develop world knowledge and problem-solving strategies that can be efficiently reused in a new task.
Who are these people?
Finally, I want to give a shout out to the awesome oaks who actually made all of this. OpenAI is a non-profit artificial intelligence research company. OpenAI also stayed true to their name and made the whole thing open source. That means that everybody can peek under the hood if they want to. I like this for much of the same reasons I just mentioned. The fact that more people can have a look at it and try to improve, is a huge advantage.
They state their mission as:
Our mission is to build safe AI, and ensure AI’s benefits are as widely and evenly distributed as possible.