So these AI Agents talks where fun. We have seen a lot of use of AI for competitive games like Go, Chess or Starcraft. But what about collaboration with other AI agents or even humans?
The first talk was about using Overcooked to research collaborative agents collaborating with humans motivated by robotic research where this problem is common. So they found how using optimal self trained agents work suboptimal together with humans because the assume humans do optimal actions (which they don’t).
The second talk was about using Bleeding Edge to train AI agents to teamwork together to defeat the other AI team. And extending that to the more challenging task of playing together with humans. This has a lot of potential for accessibility and playing/learning with bots.
The third talk was more about the help AI agents can provide for game design, more specifically content generation and testing. The question asked was, how can you generate content for games with machine learning without training data? An example of their research they have made public is rlbrush.app for creating simple socoban levels. The other part was about using AI agents to test level about playability and navigation bugs (a practical tool would be their Glitch Finder).
I really like how they give a good overview about their research topics and explain it without needing a deep detailed knowledge about a topic. The question are all really good too.
Day 2 started with an interesting session about Computational Creativity. Mark Riedl talked about using AI for storytelling and creativity in games. I always thought an AI story generator makes too much sense.
This was a good overview about the problem, the challenges, state of the art and future opportunities through the example of text based interactive story games.
The combination of ML work is what really excites me. Thing like using GPT-3 to produce story elements in text form, other algorithms to take character descriptions therein and form photorealistic characters from them and yet another algorithm to produce good voice acting.
just listening to the fireside chat with Peter Lee and Kareem Choudry
Going to be a bit of a list, I’ll keep adding to it
Alot of these things are ‘maybe in the future’ kind of things just fyi
They said that they are working heavily on using machine learning and deep reinforcement to make realistic AI NPC characters.
Also they are working on AI based content generation, and Kareem Choudry, who is the head of gaming cloud at microsoft said he is (paraphrasing) blown away by what the systems being developed in microsoft research can do
An area they are researching currently is making AI powered NPCs have ‘common sense’, making them more coherent over longer time periods
From how they are talking about AI/reinforcement learning in the gaming space I expect HEAVY investment into this, they are saying that gaming has more hours of engagement than many other types of media. (this is paraphrasing and summarizing) And that you get much more granular responses from players playing a game. Microsoft has been trying to get a social network for a long time, just for the data, so it can be used in AI research, I think they see gaming as the next untapped well of data.
looking at using AI to heighten player experiences, for instance making game moments scarier, (I guess timing jump scares and things like that?)
now talking about using AI/machine learning for moderation and matchmaking people in the best possible way to ensure a good experience
possibly doing real time chat translation for in game use (maybe)
Project Natick was inspired by xbox! bc xbox has a problem of having sudden unexpected spikes in demand in unknown regions around the world, and microsoft research came up with the idea of underwater data centres to be able to rapidly deploy date centres close to population centres
Kareem choudry now talking about his role at xbox
talking about xcloud
talks about working on reducing latency, increasing framerate and responsiveness for xcloud
They also did a Q&A but there wasnt much interesting in it
Something important to point out because I guarantee this will be used in games in the future: Deep Fakes.
Unless you follow machine learning a bit you might now know what this even is. A neural network is trained to render a photorealistic rendition of a human face and it can map that onto just about any video in place of a target’s face. It can be done at a quality that flies right over the uncanny valley and is 100% convincing. It can also be driven/animated in real time in most cases. This is a good example of how good it can look
The better results come from more source image/video training, which is great if game devs want to license the likeness of an actor without actually paying to have them do pricey performance capture. This works in game scenarios too:
This could easily end up a radically cheaper, faster, easier, simpler way to get realistic character faces into games and I bet devs start exploring it in the next few yrs.
