Thank you for the explanations :)

Actualy, I'm not realy sure I understand why japanese phantom-go is unbounded. 
Can you explain what is the difference with chinese rules ?



________________________________
De : Olivier Teytaud <[email protected]>
À : computer-go <[email protected]>
Envoyé le : Lun 13 septembre 2010, 14h 09min 31s
Objet : Re: [Computer-go] Re : phantom go, MCTS & decidability



In the genreral case of unbounded games, the question is undecidable. But you 
wonder if it is still the case for phantom go ?

For phantom-go with chinese rules, it is decidable by bounded horizon as you 
have guessed; and yes, for phantom-go with japanese
rules, I have no idea. The key point in my email is that, contrarily to usually 
published results, I consider the question of the probability of winning >=c, 
and not the question of the probability 1. This makes the question much harder, 
and in particular it is, in the general case, undecidable, even with just 2 
players.

 
What prevents a brute force approach is the fact that a strategy cannot be 
described finitely because of the unboundedness ?
>

Yes! The paper by Bob is impressive for that - it shows that a finite game (in 
the sense of a finite state space) can be undecidable (provided that the 
horizon 
is unbounded). But the model in Bob Hearn's results is the existence of a 
strategy for winning with proba 1,
which is not, in my humble opinion, a good model - such a question is not 
useful 
for playing optimally, whereas estimating the winning probability is useful for 
playing (approximately) optimally.

Strategies are answers to arbitrarily long sequences of observations. Therefore 
there are infinitely strategies, even if we consider only
deterministic strategies; and no solving by writing the matrix of the game. 
Using Madani et al, we can show that without more assumptions on the game, this 
is indeed undecidable.


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

Reply via email to