It looks like an interesting problem.
Just to be sure I understand correctly : 
In the genreral case of unbounded games, the question is undecidable. But you 
wonder if it is still the case for phantom go ?
What prevents a brute force approach is the fact that a strategy cannot be 
described finitely because of the unboundedness ?



________________________________
De : Olivier Teytaud <[email protected]>
À : computer-go <[email protected]>
Envoyé le : Lun 13 septembre 2010, 13h 38min 53s
Objet : [Computer-go]  phantom go, MCTS & decidability



Thanks Steve for your answer.

1) One-player randomized games (paper by Madani et al)
For the paper by Madani et al, they work on unobservable problems; there is a 
graph, some transitions are labelled with actions (e.g. just binary actions), 
and some others are labelled with probabilities (rational probabilities are ok, 
so there's no problem for encoding this). The instance also contains a rational 
number c, and the question is "is there a sequence of actions leading to 
acceptance with probability at least c". The undecidability result is then a 
classical undecidability result, by reduction to the halting problem for Turing 
machines.

2) Two-player games
For the extension to two players non-stochastic games, there's no more 
randomness. Unobservability is still ok.
The point is that in non-fully observable games, even when they are 
deterministic, there is a probability c of winning for an optimal stochastic 
strategy S at the worst case on the opponent's stochastic strategies S'. This 
is 
well defined, and so the question "given graph G (equipped with decision nodes 
for player 1 and decision nodes for player 2), is c bigger than d ?" makes 
sense, and is undecidable.

Importantly, approximate answers are also undecidable.

(nb: my email address is [email protected], [email protected], 
[email protected], all of them leading to the same email account)



could you be a bit more precise with your definition of the language you'd like 
to show is undecidable? try to phrase it as a decision question that results in 
a language of strings over some finite alphabet, and then maybe you can get 
somewhere, with a reduction, for instance.
>
>
>i'm not sure how you intend to define undecidability in the face of 
>randomness, 
>or even what your question is really asking.
>
>
>a natural way to use computability to include the flavor of randomness is to 
>bound the number of bad cases. i.e., to imagine that something is happening 
>either uniformly at random or with a particular distribution, and to bound the 
>probability in the uniform case is just to bound the number of bad cases in 
>the 
>deterministic case. but you have to be careful about how you state this.
>
>
>it could be a fairly straightforward undecidability result, but it depends 
>upon 
>the question that you're asking.
>
>
>PS i tried to email you directly instead, but apparently your return email 
>address is invalid...
>
>
>s. 
>_______________________________________________
>Computer-go mailing list
>[email protected]
>http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
>


-- 
=========================================================
Olivier Teytaud -- [email protected]
TAO, LRI, UMR 8623(CNRS - Universite Paris-Sud),
bat 490 Universite Paris-Sud F-91405 Orsay Cedex France http://0z.fr/EJm0g
(one of the 56.5 % of french who did not vote for Sarkozy in 2007)





-- 
=========================================================
Olivier Teytaud -- [email protected]
TAO, LRI, UMR 8623(CNRS - Universite Paris-Sud),
bat 490 Universite Paris-Sud F-91405 Orsay Cedex France http://0z.fr/EJm0g
(one of the 56.5 % of french who did not vote for Sarkozy in 2007)


      
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to