Send News. Want a reply? Read this. More in the FAQ.   News Forum - All Forums - Mobile - PDA - RSS Headlines  RSS Headlines   Twitter  Twitter
Customize
User Settings
Styles:
LAN Parties
Upcoming one-time events:

Regularly scheduled events

Out of the Blue

The new phonebooks are here, the new phonebooks are here!!!

Diabolical Links: Thanks Ant and Mike Martinez.
Play: Bowja the Ninja 2.
Science: Using Causality to Solve the Puzzle of Quantum Spacetime. Thanks j.c.f.
Media: The Website Is Down. NSFW. Thanks The Flying Penguin.
10 Best Movie Car Chase Scenes.
Geek Escher Cube Puzzle.

View
14 Replies. 1 pages. Viewing page 1.
< Newer [ 1 ] Older >

14. Re: No subject Jun 30, 2008, 16:58 Fang
 
So, if each user is a simulator training your A.I., how does it remove the processing burden from the end user?

It's removing the higher level planning from the end user. Today's games don't have very much of it, because of CPU limitations.

The end user is still running the physics simulator (hopefully running much faster on the graphics card) and managing the game world. The point of this isn't to free up CPU cycles on the end user, because very little of it is being used for higher-order planning. It's to provide something that isn't being done (or only in a very limited fashion) right now.

You are now contradicting yourself from earlier, this is no-longer real time with regards to gameplay.

You only think this because you don't understand how reinforcement learning works.

The model-based controller in the helicopter is running in real time. In the same way, the software agents in the end-users game would be running in real time. Don't confuse the learning process with action process. The action process response from the central AI server (the "here are my observation what should I do") can occur very quickly, especially since the goal isn't to give a truly optimal answer, but instead a best guess given the current learning level. It's the equivalent of a database query. The learning process (which took a day with the helicopter example) is about generating a higher order plan that does something truly intelligent, like fly a helicopter upside, or respond to an ever changing tactical situation. This occurs in the central AI server farm, so it doesn't have to be done for each user. This is where you get all of your processing savings compared to the current inefficient design.

At first, the game's agents probably won't be that smart. But with 100,000 simulators, the central AI server would probably start giving very intelligent actions pretty quickly.

Kind of like the Folding/SETI@Home programs. The more systems you have, the faster you can solve the problem.

Think of it this way. Do you remember the Sony AIBO? The little robotic dog they made? Each dog, when it comes out of the box has a limited controller that is passable for what people expect. Kind of like today's games. Now if each of these dogs were networked with a central server, the central server can issue the policies for how it thinks the controller should behave. This occurs very quickly and with little overhead. It will take time for the actions to truly become intelligent, but the dogs will work, and the games will play. If you had 100,000 dogs working at the same time, it would happen very quickly.

Of course, there is a exploration vs exploitation tradeoff, but I won't get into that.

Since most single player games have already been designed to run with a multiplayer server, there really isn't that much more overhead you are adding to the user client. And unlike in a multiplayer game, you can continue to let the client handle the gameworld physics.

 
Reply Quote Edit Delete Report
 
13. Re: No subject Jun 30, 2008, 08:30 Enahs
 
So, if each user is a simulator training your A.I., how does it remove the processing burden from the end user? You have just added more overhead on the CPU with regards to communicating the information and bookkeeping with the master A.I. server.


For example, how much processing power do you think it takes to have a computer fly a real helicopter upside down, and perform other acrobatic maneuvers that no human could possibly perform? It has been done, and its a lot less than you'd expect. It does take a day before hand to train it though in a simulator.

A proper design would have the centralized server (more like a server farm, no need to use expensive supercomputers like enahs suggests) taking inputs from all users and combining it to create a robust policy for specific observations that the client-based agents generate. Each additional user is actually a simulator to train the centralized AI. And a policy query to the central server farm would just be a database query.

You are now contradicting yourself from earlier, this is no-longer real time with regards to gameplay.


You are not talking about running the A.I. on a central server. You are talking about sending gameplay information and learning, and the updating A.I. and then updating the client.

That is different from the A.I. running remotely on a computer with more processing power. But you have not freed up any processing power from the end user.


Completely different thing then you talking about earlier.




Alternating Logo (GreaseMonkey script):
http://www.ualr.edu/szsullivan/scripts_/BluesNewslogo.user.js
 
Avatar 15513
 
I am free of all prejudice. I hate everyone equally.
- W. C. Fields
Reply Quote Edit Delete Report
 
12. Re: No subject Jun 29, 2008, 22:34 Fang
 
And you prove my point about why you can't just hire some programmers to try to do this, and you need computer scientists.

You're comparing apples to oranges. You're trying to just scale old AI algorithms and assume the implementers would design it stupidly. That 10% would be mostly redundant for each user.

For example, how much processing power do you think it takes to have a computer fly a real helicopter upside down, and perform other acrobatic maneuvers that no human could possibly perform? It has been done, and its a lot less than you'd expect. It does take a day before hand to train it though in a simulator.

A proper design would have the centralized server (more like a server farm, no need to use expensive supercomputers like enahs suggests) taking inputs from all users and combining it to create a robust policy for specific observations that the client-based agents generate. Each additional user is actually a simulator to train the centralized AI. And a policy query to the central server farm would just be a database query.

As a computer scientist, this scenario would be amazing for me. Can you imagine having access to 100,000 simulators to train your algorithm? Eegads, right now I can only run 1 simulator at a time.

So in fact, more users will greatly increase the centralized AI's available processing power. Of course, it would have to be able to integrate the data these users generate, but the necessary server farm would be on the order of magnitude of a web server farm handling the same number of concurrent users.

If you don't quite understand this, I'd recommend taking a look into POMDPs and Policy Search by Dynamic Programming.

 
Reply Quote Edit Delete Report
 
11. No subject Jun 29, 2008, 19:52 Mr. Tact
 
I need a new phonebook. My current one is from 2006.  
Truth is brutal. Prepare for pain.
Reply Quote Edit Delete Report
 
10. Re: No subject Jun 29, 2008, 17:57 Enahs
 
You are talking about things that are not done in real time.

You are saying the ~10% or so processing power the A.I. currently gets on home CPU's is not enough. So, theoretically, if it is 10% and they double the power to 20%, your CPU could handle the client for 5 people then, ignoring the overhead of multiple things going. So, for 100,000 to play the single player game, they would require 20,000 of your theoretical home CPU's. The top supercomputer right now only has 6,480 CPU's.


Yes, maybe you can do centralized learning as you say, but it will still have to do calculations for each individual user.

It would require a freaking huge amount of processing power.


Alternating Logo (GreaseMonkey script):
http://www.ualr.edu/szsullivan/scripts_/BluesNewslogo.user.js


This comment was edited on Jun 29, 18:16.
 
Avatar 15513
 
I am free of all prejudice. I hate everyone equally.
- W. C. Fields
Reply Quote Edit Delete Report
 
9. Re: No subject Jun 29, 2008, 16:38 Fang
 
The amount of processing power required for that would be huge.

Not if you design it correctly. It's possible to use linear dynamic programming to apply machine learning techniques.

One example of a leap forward in applied AI algorithms is with Bayesian spam filters. They can update their models very quickly and they work reasonably well once trained. Much better than the spam filters that came before them.

That's just an example, and there are plenty of cutting edge AI techniques that will scale very well, especially since the point of this would be to run a centralized learning system. No need to relearn for each user.

Obviously, you can't just hire more programmers to try to implement it; you'll have to hire true computer scientists to implement a next-gen AI solution.

 
Reply Quote Edit Delete Report
 
8. Re: No subject Jun 29, 2008, 12:22 Enahs
 
The amount of processing power required for that would be huge.



Alternating Logo (GreaseMonkey script):
http://www.ualr.edu/szsullivan/scripts_/BluesNewslogo.user.js
 
Avatar 15513
 
I am free of all prejudice. I hate everyone equally.
- W. C. Fields
Reply Quote Edit Delete Report
 
7. Re: No subject Jun 29, 2008, 00:48 Fang
 
Worst idea ever?

I'm sure that's what people said about paying $15/month to play a MMO. Who would want to do that? That's paying $180/year to just play one game! Well, we know how well that worked out for Blizzard. Aren't they going bankrupt pretty soon?

If a developer was just running the same old AI code on a server, that would be silly. The point of this would be to be able to utilize another machine to be able to do next-gen AI. Currently, the user's CPU is getting hammered trying to handle the graphics and physics. Right now, there's a tradeoff on how much CPU you can dedicate to the AI without impacting framerate.

What if we could apply machine learning techniques to make the computer much more life-like? Or do you just prefer playing against the same dumb enemies?

Maybe paying $20 for a year subscription to a single player game with jaw-dropping AI (or however the business model turns out) is not something you would be interested in. To be honest, I'm not that interested in paying $15/month for a multiplayer game, no matter how massive it is. So I don't. However, many people do pay for WoW, and I'm sure there's a business model that would work for a subscription to a game AI server.

 
Reply Quote Edit Delete Report
 
6. Re: No subject Jun 28, 2008, 21:45 Jerykk
 
What if a developer created their single player game to run all their AI on developer run servers?

Worst idea ever?

I'd be curious to see the impact on sales to a developer like Crysis.

Much, much lower, I'm guessing.

 
Avatar 20715
 
Reply Quote Edit Delete Report
 
5. Re: No subject Jun 28, 2008, 15:42 Enahs
 
The Website Is Down is great!

Alternating Logo (GreaseMonkey script):
http://www.ualr.edu/szsullivan/scripts_/BluesNewslogo.user.js
 
Avatar 15513
 
I am free of all prejudice. I hate everyone equally.
- W. C. Fields
Reply Quote Edit Delete Report
 
4. Re: No subject Jun 28, 2008, 15:21 Enahs
 
I would put Bullitt at number one. Speaking off, do not forget to check out today's metaversa section with the car chase scene from Bullitt as well as tracking it "live" on google maps.



Alternating Logo (GreaseMonkey script):
http://www.ualr.edu/szsullivan/scripts_/BluesNewslogo.user.js
 
Avatar 15513
 
I am free of all prejudice. I hate everyone equally.
- W. C. Fields
Reply Quote Edit Delete Report
 
3. Re: No subject Jun 28, 2008, 13:33 Fang
 
I wonder how much bandwidth is needed to stream a gameplay screen (ie. DirectX output) to a user. Probably too much.

What if a developer created their single player game to run all their AI on developer run servers? It would put a serious cramp on piracy, as long as the AI servers don't get out into the wild.

Of course, users would be aware that that if the developer/publisher goes under, then the game no longer works. The company should probably sell it as a subscription service for a specified time period and price it accordingly.

Hrmm, this would kind of be like an MMO without the multiplayer aspect of it. I'd be curious to see the impact on sales to a developer like Crysis.

 
Reply Quote Edit Delete Report
 
2. Re: No subject Jun 28, 2008, 13:20 Syntarsus
 
"I'm in print! Things are going to start happening to me NOW."

And as a connoisseur of car chase scenes, that top 10 list is one of the better ones I've run across. I'd toss out French Connection on the technicality of it being a car chasing a commuter train, which isn't even running from the car. I'd also lose Against All Odds (it's a race scene) and Smokey (not that good), and put in The Seven-Ups, What's Up Doc, and The Junkman. Then you'd have a pretty kick-ass car chase scene list.

 
Reply Quote Edit Delete Report
 
1. No subject Jun 28, 2008, 10:50 xXBatmanXx
 
hehehehehe Yeah, mine are sitting right here near me, still in the bag......we will bring them to Wisc to shoot them, then burn them. BURN THE WITCH!

You got a great body, but your record collection sucks....
BN 360 Scoreboard:
http://www.mygamercard.net/clboard.php?id=GW57kfq7
 
Avatar 10714
 
In this present crisis, government is not the solution to our problem; government is the problem. / Few men have virtue enough to withstand the highest bidder.
Playing: New dad
Reply Quote Edit Delete Report
 
14 Replies. 1 pages. Viewing page 1.
< Newer [ 1 ] Older >


footer

.. .. ..

Blue's News logo