Ai Dreams Forum
Games => Gaming => Topic started by: Freddy on July 04, 2017, 08:37:38 pm
-
Looks as if Unity are having a pop at this. No info posted, just a video.
https://www.youtube.com/watch?v=fiQsmdwEGT8
-
This is all very good for playing games but what practical use does it have? I've seen a few of these AIs playing Super Mario for 6 hours until they become good but let's imagine we have a surgical AI. Would we want it chopping up people randomly for 6 hours before it learned how to use a scalpel?
-
This is a good point. I suppose it's one of those things where you don't think of a use for it until after you have played with it a while. Often distractions can lead to useful things.
And there is the just for the heck of it angle too - it's an interesting programming challenge.
It would give us smarter opponents in computer games though, perhaps.
-
This is all very good for playing games but what practical use does it have? I've seen a few of these AIs playing Super Mario for 6 hours until they become good but let's imagine we have a surgical AI. Would we want it chopping up people randomly for 6 hours before it learned how to use a scalpel?
That's what medical students do, either with cadavers or in simulations.
https://zygotebody.com/
-
There are plenty of tasks that can be learned through trial-and-error. Training autopilots through flight simulations, for instance, or programming, social interaction, design, war strategies. It's not like babies don't try things a hundred times before getting it right. You could also teach AI a task another way and then use reinforcement learning to fine-tune the details with experience, although the scientists have yet to reach that wisdom.
war strategies (https://www.youtube.com/watch?v=NHWjlCaIrQo)
-
Nice one. Funny but our Smart Phones are probably more powerful than that WOPR computer from that era.
Reinforcement is one of the best teachers out there. (IMHO)
-
Good points Don. War Games was a brilliant film wasn't it, and it still is.
-
This is all very good for playing games but what practical use does it have? I've seen a few of these AIs playing Super Mario for 6 hours until they become good but let's imagine we have a surgical AI. Would we want it chopping up people randomly for 6 hours before it learned how to use a scalpel?
That's how human surgeons learnt to do what they do. There's a never ending supply of cadavers out there to help these things learn. They tend to be more accurate that humans. No worries about if they had too much coffee or distractions during surgery. Pretty much everything we have today is a result of trial and error.
-
Good points Calhoone.
No one ever set out saying, "I'm going to discover __blah...blah__ today!"
The majority of Discoveries were actually the result of an accident or failure. Some other discoveries were the result of applying what was learned from these failures until...Eureka! (light bulb goes off...or rather on in this case!) O0
-
This is all very good for playing games but what practical use does it have? I've seen a few of these AIs playing Super Mario for 6 hours until they become good but let's imagine we have a surgical AI. Would we want it chopping up people randomly for 6 hours before it learned how to use a scalpel?
Games are simulations, but well they are meant for entertainment. They can be used to try out stuff easily, like games were quite often the driving force of new technologies (e.g. VR). But more important, real world applications can benefit from transfer learning. In a real world, it is not really feasible to let a car crash over and over again into different props thousands of times. As an example, GTA IV is capable of outperforming real world traffic video footage (http://download.visinf.tu-darmstadt.de/data/from_games/) (CamVid Dataset). And this is what Demis Hassabis (DeepMind) said (https://www.youtube.com/watch?v=rbsqaJwpu6A&feature=youtu.be&t=9m21s) about games.
To append to Unity's video, they started to publish some blog posts (https://blogs.unity3d.com/2017/06/26/unity-ai-themed-blog-entries/) about reinforcement learning.
-
Actually I have done the opposite, I have discovered hundreds of human intelligence (and everything else) instructions by pure want, - search and ye will find. I have huge enthusiasm to make it happen.
The AIs will explode even more the same way, and yes by random accidents too. You need to create a database/knowledgebase - and you will, by agenda and desire, find and landmark many huge discoveries, and projects.