|
Post by Hannes Vilhjalmsson on Feb 12, 2009 16:36:01 GMT -5
This week you have been reading "Simulation-Based Approach to General Game Playing" by Hilmar Finnson and Yngvi Björnsson. This reading gives you a chance to see what is going on inside the worldchampion in general game playing. This coming Tuesday, Yngvi will be a guest in our discussion class, so this is your opportunity to learn from the games master himself Post your discussion questions here by Sunday night (the very latest, preferably by Friday night).
|
|
|
Post by arnists on Feb 13, 2009 4:45:58 GMT -5
(I'm assuming the questions/comments are directed @ Yngvi) For anyone interested in the actual games played, a list can be found @ this pageYou can look up the agents used and actual games played by using the navigation bar. I suppose it's not possible to evaluate the complexity of a game automatically from it's description due to emergent games being hard to analyze. I'm intrigued by the utter lack of epistemology in CadiaPlayer. One of the strongest factor of human-level intelligence is the ability to abstract, process the abstraction and form a higher-level strategy from it (sacrificing a piece in chess as an example). Being epistemologically unaware has the disadvantage of only walking the path you are on without being aware where you were. I found it intriguing to browse through actual games played and found an instance of the 8-puzzle where CadiaPlayer shifted the same piece back and forth several times ( game). Do you have an example of a game that CadiaPlayer has a problem playing (I mean where other approaches obliterate it)? What is your next move?
|
|
|
Post by Helgi Páll Helgason on Feb 13, 2009 8:19:36 GMT -5
Can you give us examples of applications for general game playing outside of playing strictly structured games like Checkers and Chess? Can GGP be used for pseudo-continuous environments with loose structure? For example, a first person shooter game environment with two players trying to destroy each other.
Have you considered adding a high level module to CADIAPlayer that monitors the internal behavior from "above" and can change parameters and behaviors of the player when it sees opportunities for improvement? A sort of module that would reason about and affect the behavior of the underlying systems as a whole through time.
|
|
|
Post by Christian Zehetmayer on Feb 13, 2009 14:09:41 GMT -5
1) How much influence has the CPU power or parallelization to win a GGP competition?
2) How much effort did you (Yngvi) use to develop CadiaPlayer or respectively to win the AAAI GGP competition?
|
|
|
Post by Richard Ottó O'Brien on Feb 14, 2009 10:10:22 GMT -5
I was quite fascinated by the paper, in that I found it was a very interesting reading. It fit perfectly to what I for many years thought was the definition of an A.I system. As I have always thought that an almost autonomous system which gets some rules and limits about its environment and then "by itself" figures out a way to solve a problem or in this case play and win a game.
1) Is there a possibility to see the CadiaPlayer play against another player? or play against it? I would love to see an entire match of some game. 2) What is the results from creating such GGP systems? Will these be implemented into some commercial games, which actually play a fixed game rather then general games? or is the goal to discovered better working algorithms?
|
|
|
Post by halldorrh05 on Feb 14, 2009 11:41:02 GMT -5
1) How exactly does cadiaplayer "learn" to play the game? 2) How hard would it be to make cadiaplayer "cooperative" in order to have meaningfull teamwork?
|
|
|
Post by Birna Íris on Feb 15, 2009 6:21:13 GMT -5
1) Any thoughts of other domains (outside the game domain) this work could be expanded in?
2) I didn't understand well enough how using the UCT algorithm bypasses the need for a heuristic evaluation function. Would it be possible to combine the two?
|
|
|
Post by Jón Trausti on Feb 15, 2009 11:15:10 GMT -5
1) In chess, is there any known AI that beats all human players? If so, can you give me a glimpse at what method it uses?
2) If chess could be played perfectly, as in, if there's always one perfect move from any state, wouldn't the game end up being drawn?
|
|
|
Post by Hjalti Kolbeinsson on Feb 15, 2009 12:16:49 GMT -5
1) You have already won the AAAI GGP competition 2 years in a row, and I assume you will try to win the third time. How have you altered the Cadia-player this year?
2) According to the paper the cadia-player ignores other players in multiplayer games where opponents are more than one. Do you have any ideas on what kind of algorithm to use if you intend to change this?
|
|
|
Post by hordur08 on Feb 15, 2009 12:52:42 GMT -5
1) You say in your article that better actions are selected more often than suboptimal ones. I didn´t really get it how CADIAplayer chooses which actions to take. Could you explain further how CADIAplayer selects nodes to expand.
2) Are there any types of games that CADIA player is especially good at and are there any types of games that it is poor at?
|
|
|
Post by olafurgi on Feb 15, 2009 14:50:11 GMT -5
What are your next steps in improving the CadiaPlayer ?
Is there any specific reason (performance/readability) for your choice of programming-language for the agent ?
|
|
|
Post by Jon Gisli Egilsson on Feb 15, 2009 15:58:20 GMT -5
1) Can we see any kind of GGP agents in things we use on a daily basis?
2) Why did you have to run experiments to see if an agent would beat another identical agent given more time to think? Isn't that obvious?
|
|
|
Post by Alfreð Már Alfreðsson on Feb 15, 2009 17:12:18 GMT -5
1. Can you explain better the method for single-agent games the engine uses, Memory Enhanced IDA.
2. You know how they intend to improve CadiaPlayer for the next GPP event?
|
|
|
Post by Bjarni Gunnarsson on Feb 15, 2009 17:13:16 GMT -5
Would it be possible to split the CadiaPlayer's work over multible computers or would the network latency/bandwith kill the deal?
Has the CadiaPlayer been used to solve problems other than the GGP?
|
|
|
Post by Snorri Jónsson on Feb 15, 2009 17:13:52 GMT -5
1. When Deep blue whon a match it was because of a manual adding game-specific knowledge. How far along are agents in using learning to defeat a human in a game of chess?
2. CadiaPlayer won the AAAI GGP competition again in 2008 was it because of improvements? if so what improvements and what is the plan for next year?
|
|