|
Post by Hannes Vilhjalmsson on Feb 8, 2011 7:36:06 GMT -5
We started last week talking about game playing and we will continue with that theme for the next couple of lectures. It turns out that we have one of the greatest experts on autonomous game players among us: Yngvi Björnsson. He was on the team that solved Checkers and with his students at HR, he is a double world-champion in general gameplaying. Yngvi will be our guest speaker on Monday, so make sure to use this opportunity to learn what you've always wanted to know about game playing computers. The reading for the discussion is the following paper: Yngvi Björnsson and Hilmar Finnsson. CadiaPlayer: A Simulation-Based General Game Player. IEEE Transactions on Computational Intelligence and AI in Games, 1(1):4–15, 2009.The questions you post below will be given to Yngvi on Monday morning for review. Fire away :-)
|
|
|
Post by carmine on Feb 10, 2011 13:36:06 GMT -5
- What do you mean with "balance between exploring and exploting"? - Is your approach useful in other complicated problems that are not about "games"?
|
|
|
Post by Hrafn J. Geirsson on Feb 12, 2011 10:34:27 GMT -5
1) How does the CADIA Player go about cooperating with other players ?
2) How does simulation based search work, and why is it so easy to parallelize ?
|
|
|
Post by lorenzo on Feb 12, 2011 12:10:37 GMT -5
- In the last paragraph of the paper the author introduce same future work on this tecnology so i want to know if they start to work on it and if they have already some result.
-How much time the UTC algorithm recognize that is taking the wrong decision and change the rule for take the correct one.
|
|
|
Post by Eiríkur Fannar Torfason on Feb 12, 2011 15:00:23 GMT -5
1. The paper and the empirical studies seem to focus solely on either turn-based games or puzzle (single agent) games. Is CadiaPlayer capable of playing real-time games and if so, how well does it perform?
2. How much of an effect does the level of branching in the game tree have on the performance of simulation based agents?
|
|
|
Post by kristjan on Feb 13, 2011 10:28:03 GMT -5
For the GGP competition what are some of the worst performances you know of? Did any of the agents get zero wins in any game? And how big is the number of participants in the competitions?
|
|
|
Post by thorsteinnth on Feb 13, 2011 10:37:28 GMT -5
Í greininni er sagt að til að spara minni þá megi eyða öllum nóðum úr MC leitartréinu sem eru fyrir ofan nóðuna sem maður er í. Mætti í raun ekki eyða öllu hinu hluttréinu? Segjum að við erum í nóðunni A, og ferðumst til vinstri, mætti þá ekki eyða öllu hægra hluttréinu úr minni?
Þegar sagt er að agent geti lært eitthvað spil, er þá verið að meina að hann fái sett af reglum og læri hann þá instant, eða þarf hann að spila nokkra leiki til að geta lært?
Hvaða áhrif hefði það á "lærdóm" hjá agent ef hann myndi allt í einu þurfa að spila aðra útgáfu af spili sem hann hefur lært áður, t.d. mismunandi útgáfur af póker eða eitthvað þannig?
|
|
|
Post by Ásgeir Jónasson on Feb 13, 2011 11:17:11 GMT -5
Does the CADIAPLAYER try to analyze how the other players might be choosing their moves, based on observation, or is there no time for such calculations ?
Does the CADIAPLAYER's simulation method always outperform the traditional, minimax-based game tree-search, or is it dependant on the time and space constraints. For example, does the minimax method outperform CADIAPLAYER when there is very little time to think or memory to use ?
|
|
|
Post by Þorgeir Karlsson on Feb 13, 2011 11:22:59 GMT -5
-What kinds of implications could this research have on other fields of computer science? e.g. could this lead to an intelligent agent capable of making good business or investment decisions?
-Has chess not been used at the GGP competitions?
|
|
|
Post by Helgi Siemsen Sigurðarson on Feb 13, 2011 13:35:25 GMT -5
1) "There are fundamental differences between that method and the one we propose here for search control, the main one being that data are propagated differently: in our method from the top-level tree down to the playout phase, whereas in the aforementioned method, it is the other way around." what does this mean ?
2) what are stopping the coevolution approach and the logic program approach mentioned in chapter 6 from being developed ?
|
|
|
Post by Jökull Jóhannsson on Feb 13, 2011 14:05:03 GMT -5
1) Is there any mone in making ai for turn based games. 2)What is the main diffrense on making ai for turned based games and real time games ?
|
|
|
Post by helgil08 on Feb 13, 2011 14:59:59 GMT -5
1) Is there any way to make use of this for say, a "normal" computer game instead of having to write a new AI for each game.
2) Just how complex is all this. When working with this, what is most likely to go wrong.
|
|
|
Post by finnur on Feb 13, 2011 15:16:52 GMT -5
-Does the CadiaPlayer analyze the opponents moves to determine what he is most likely to do next?
-If not, would it make him better to do so?
|
|
|
Post by jonfs09 on Feb 13, 2011 16:16:25 GMT -5
What exactly is this monte carlo game tree?
In this system Pell’s METAGAME, what is the difference between the CLUNEPLAYER and FLUXPLAYER and how do they work?
|
|
|
Post by grimurtomasson on Feb 13, 2011 16:21:19 GMT -5
Is CADIA Player still focusing on Monte Carlo tree search and UCT?
Is the A* based solver still being developed?
Were parts of CADIA Player ever ported to a GPU?
On what is your current research focused (Yngvi)?
|
|