You know that'd be really interesting as spectacle (in my opinion). You'd pick a game that neither company has worked on and pit them against one another.
From a research standpoint though, it'd be a waste of time.
It wouldn't be a waste of research on that game. But for the broader community it might effectively be. I think it'd be a great way to setup a competition among grad students though.
People mentioned in the article seem to doubt about applicability in the real world. If nothing else, I think it brings video game development closer to being an automated process. A game could simulate various conditions and design an infinite number of complex environments. Am I way off?
I saw an interesting talk about proc gen which was basically saying a lot of studios are moving from hand crafting to procedural tools. For example if you want to fill a park with trees you don't place them by hand, you just specify the density and type of trees and let the computer place them. The advantage is that if you want to reshape the park later you can just click and drag. Then when you are happy you go though and hand craft more content on top.
They were saying there is a critical mass where procedural tools feed into procedural tools and a lot of the pipeline is automated, then it becomes extremely cheap to make more content.
So back to your point: I think this is where we will see AI more in game dev, it will be as smart tools to empower creators to make stuff more easily. I don't think the AI will make the game, I think it will assist the humans with the creation of it.
I would see this being used more for automating many aspects of testing games, then down the line - level design. But for automating game development, you would require creativity - unless you're into sudoku style games.
An important detail is missing from the article (or perhaps I missed it). The AI had an advantage unrelated to intelligence. It's "plugged" into the game in a way that gives it a big advantage. Better reaction time, better aim, technical capabilities that a human player with "primitive" input devices and looking at the game through a low res window can't get close to.
Unlike a game of chess where the mind is basically the only one at work, this time the "body" might have had just as much of an input. Even regular bots were good enough for most players anyway :).
> at close range, FTW excelled in combat, in part because its reaction time was half that of a human's, and in part because its accuracy was 80 percent compared to the humans' 50 percent. [0]
I see now that reaction time was compensated for. No info on accuracy.
People said the same thing about DeepMind's SC2 wins. Its actions per minute spiked to inhuman levels briefly and its efficiency and precision of control was nearly perfect. These types of criticisms completely miss the point.
People were attempting to build StarCraft agents for 10 years. 50000 actions per minute, complete vision of the entire map, predefined scripted build orders, restrictions on human opponents - nothing worked. A moderately competent player would destroy those old bots in 10 out of 10 games, it wasn't close.
These games have so much more to them than reaction time and dexterity, these new agents are doing something that simply wasn't possible before.
That's because that AI would always have some exploitable flaws that people pick up on really quickly. I wouldn't be sure that DeepMind wouldn't fall into the same problem, it would simply take longer to get there.
> These games have so much more to them than reaction time and dexterity
Sure but they're still an important part of the end result which most people don't consider. For many the title just says "AI beats human mind". If you want to know which mind is better then compare minds, all else equal.
"Dumb" bots have been kicking player's butts for years now. Not because of intelligence but because everything else was not equal: great aim, great reaction time, great situational awareness. So aren't you curious how much the "intelligence" part evolved?
Imagine a 100m sprint where one of the runners has to wear army boots and backpack. Would you even bother drawing a conclusion?
The original technical paper (I posted a link elsewhere in this thread) discusses that the AI only had access to the same visual imagery a human player can access
A benchmark for me would be a game that can handle a MMOPG like eve-online.
What with certain tactics in game valid - like griefing, spies etc. An AI can handle that would be not only impressive, but learn a set of skills that many would feel uncomfortable about being within the realms of AI.
After all, humans cheat, would we want that allowed by an AI. Equally an AI could stumble upon an exploit in a game and not realise it is an expliot - ie item duplication.
But this opens up a whole new aspect of gaming that AI has yet to touch that I'm aware of.
Though I feel we are still a way of from having computers with a moral-coprocessor, and still very much early days with AI. But that is certainly something that is within our lifetimes.
MMOPG's are not all reducible into an Excel spreadsheet and the part were I said "What with certain tactics in game valid - like griefing, spies etc. An AI can handle that would be not only impressive, but learn a set of skills that many would feel uncomfortable about being within the realms of AI." was the part I was discussing.
While I am a fan of DeepMind’s research (Alpha Go and Alpha Zero blew me away - I wrote a commercial Go program in the late 1970s for the Apple II so I appreciate the problem, and I used to work with DeepMind cofounder Shane Legg at Webmind), respectfully, I do have a criticism:
I would like to see them use their talent and financial backing from Alphabet to tackle the hard problems of common sense in AI and in general the science of what it will take to reach AGI.
Deep learning is brittle, as can be seen in the good science of finding adversarial attacks on trained models. I would like to see less accuracy on dev and test sets, smaller more general models, and especially not see convolutional models as the basis for vision systems in self driving cars. My intuition is that discarding information of relative positions and relative relationships of large features in an image is a bad idea for the self driving car problem.
They have really pushed progress for an old technology (reinforcement learning) and achieved fantastic results but in my experience if there are not many failures along the way, then perhaps goals are not being set high enough.
25 comments
[ 3.0 ms ] story [ 31.7 ms ] threadFrom a research standpoint though, it'd be a waste of time.
[0] https://news.ycombinator.com/item?id=20055162
They were saying there is a critical mass where procedural tools feed into procedural tools and a lot of the pipeline is automated, then it becomes extremely cheap to make more content.
So back to your point: I think this is where we will see AI more in game dev, it will be as smart tools to empower creators to make stuff more easily. I don't think the AI will make the game, I think it will assist the humans with the creation of it.
Here is a talk about Spiderman and how much proc gen they used there https://www.youtube.com/watch?v=4aw9uyj9MAE
Unlike a game of chess where the mind is basically the only one at work, this time the "body" might have had just as much of an input. Even regular bots were good enough for most players anyway :).
I see now that reaction time was compensated for. No info on accuracy.
[0] https://arstechnica.com/science/2019/05/googles-ai-group-mov...
People were attempting to build StarCraft agents for 10 years. 50000 actions per minute, complete vision of the entire map, predefined scripted build orders, restrictions on human opponents - nothing worked. A moderately competent player would destroy those old bots in 10 out of 10 games, it wasn't close.
These games have so much more to them than reaction time and dexterity, these new agents are doing something that simply wasn't possible before.
Sure but they're still an important part of the end result which most people don't consider. For many the title just says "AI beats human mind". If you want to know which mind is better then compare minds, all else equal.
"Dumb" bots have been kicking player's butts for years now. Not because of intelligence but because everything else was not equal: great aim, great reaction time, great situational awareness. So aren't you curious how much the "intelligence" part evolved?
Imagine a 100m sprint where one of the runners has to wear army boots and backpack. Would you even bother drawing a conclusion?
The reference is simple: SkyNet fought war, and a war is a type of multiplayer game.
Am I missing something?
[1] - https://www.popularmechanics.com/military/aviation/a26871027...
After all, humans cheat, would we want that allowed by an AI. Equally an AI could stumble upon an exploit in a game and not realise it is an expliot - ie item duplication.
But this opens up a whole new aspect of gaming that AI has yet to touch that I'm aware of.
Though I feel we are still a way of from having computers with a moral-coprocessor, and still very much early days with AI. But that is certainly something that is within our lifetimes.
I don't think they managed to connect Excel to their game environments yet.
I would like to see them use their talent and financial backing from Alphabet to tackle the hard problems of common sense in AI and in general the science of what it will take to reach AGI.
Deep learning is brittle, as can be seen in the good science of finding adversarial attacks on trained models. I would like to see less accuracy on dev and test sets, smaller more general models, and especially not see convolutional models as the basis for vision systems in self driving cars. My intuition is that discarding information of relative positions and relative relationships of large features in an image is a bad idea for the self driving car problem.
They have really pushed progress for an old technology (reinforcement learning) and achieved fantastic results but in my experience if there are not many failures along the way, then perhaps goals are not being set high enough.