Offline Simulation / Access to points (reward)

General questions and discussion

Offline Simulation / Access to points (reward)

Postby mt80 » Fri Apr 19, 2013 1:34 am

Hi all!

I have two questions:

1) is it possible to run the simulation in offline (no browser) mode? This would be a great advantage for incorperating trial-and-error methods, such as reinforcement learning!

2) is it also possible to receive messages about the success of a shot (the points)?

Cheers

Mt80
mt80
 
Posts: 1
Joined: Fri Apr 19, 2013 1:27 am

Re: Offline Simulation / Access to points (reward)

Postby admin » Fri Apr 19, 2013 11:10 am

Hi!

1) I don't see why this shouldn't be possible, but this is up to you to implement. The only information your agent will get from the server during the competition is a screenshot of the game.

2) We are currently finalising the client/server protocol and it will be possible to query the game state. But you will have to ask the server for it, the server will only answer requests and never send out unrequested messages.
We are testing the client/server version of our software at the moment and will make it available soon (probably next week).

Cheers,
Jochen
User avatar
admin
Site Admin
 
Posts: 9
Joined: Tue Mar 19, 2013 5:20 pm

Re: Offline Simulation / Access to points (reward)

Postby jasonli » Tue Jul 16, 2013 8:33 pm

One possible argument against offline simulation (with a physics engine) is that it takes away the "AI" part of the problem.

The physics engine of Angry Birds (Box2D) is publicly available. It is possible to ascertain the Box2D parameters for the blocks and recreate entire scenes. Thus, one can understand exactly the consequence of the shots, therefore reducing the problem to a breath-first-search of depth 3 or 4 (typical number of birds). I would say that this trivializes the problem, and is equivalent to opening up the internal game engine. The agent with Box2D component would therefore have an unfair advantage, and there would be no use for "AI" in this approach.
jasonli
 
Posts: 2
Joined: Tue Jul 16, 2013 8:24 pm

Re: Offline Simulation / Access to points (reward)

Postby Jochen » Wed Jul 17, 2013 12:06 pm

Hi Jason,

think about it, how many possible shots are there per bird, how long does it take to simulate a shot and how many shots can be simulated in the given time? In addition, agents only get screenshots and not the exact information, so exact simulation seems impossible. Also note that last year at least two agents used box2d and they didn't win.

We will see what works best this year, but if using box2d gives a major advantage, then we could split it into two different tracks next year.

Jochen
User avatar
Jochen
 
Posts: 86
Joined: Tue Mar 19, 2013 10:36 pm
Location: Canberra, Australia

Re: Offline Simulation / Access to points (reward)

Postby jasonli » Wed Jul 17, 2013 7:00 pm

It is not as impossible as it sounds. Box2D is a physics simulation engine, for which you can tune all available parameters. It is possible that exact parameters can be derived offline prior to the competition, and then tuned so that you speed up the simulation by several folds, so you get the exact consequences for each shot. I'd be surprised that if this is implemented correctly, it doesn't beat the naive agent. However, I still feel that this takes away the "AI" part of the problem, which is to estimate the consequences of actions without a physics engine to act as a oracle.
jasonli
 
Posts: 2
Joined: Tue Jul 16, 2013 8:24 pm


Return to Q&A

Who is online

Users browsing this forum: No registered users and 6 guests

cron