Hi there.
To my best knowledge, most people do use Gogui and gtp. This provides
interesting ways to see analysis results. But only in a "flat" way.
When the need arise to look deeper in what the bot is thinking, one
seems to need to develop his own system. It's quite hard in fact, and
lack interactive features.
Personally, when i make up a bot, and try to debug it, and to improve
it : i like very much to understand how it reaches the conclusions it
does reach. I like to be able to visualize not only the score of the
top node, but the score of any deep node.
PART 1 :
Basically, i'd like GTP (or one of it's extensions like gogui
provides) to feature a standard way for the exploration of deepnodes
states. And not only for the rootnode.
For an example, i may wonder why the program doesn't not explore a
prety obviously good branch. I'd like there to be a command for me to
do so. I'd describe the branch i want to know more about, and then
launch an analysis command for that particular branch. It could be
achieved through the concept of "path from the main node".
Analysis command would not only be able to be passed for the root
node, but for any node through a path. A path there being a list of
valid moves.
Then one would need a gui to use that easily. The only easy to use
solution that came to my mind. (and i actually implemented it, in some
dirty ways) was to duplicate the board view. So i would have two
boards. One main board that does work exactly as gogui board do. And
one VARIATION board, that is used to build up a path through clicks and
visual interaction. The VARIATION board is the one i would like to see
the result of the analysis_through_path command.
English is not my native language, so i have no idea if what i said was
clear or not. (feel free to interpet and put it in a better english, if
you think you grabbed the idea ...)
- Do you think that having a gui and protocol to help explore deeper
nodes would be usefull ? Would you like it to be integrated in an
already mature and nice environement like gogui ?
PART 2 :
Next, most bot use some kind of iterative exploration. That eventually
converge to a good result. Sometime starting with very bad
understanding of the board position. I'd like to be able to have more
(interactives) informations about those behavior too. So maybe it would
be possible to add a command to the standard genmove, so it could be
added a parameter that would set the bot for a certain number of
iteration analysis. Of course it is most usefull with what i ask for in
part1 ..
So when you play a tournament, and your bot loose in a position. You
would be able to launch gogui, set up the position. And then explore
visually what your bot think for different variations, with any sort of
depth from the effective position you want to analyse. And you'd be
able to visualize how it does converge (or diverge) from what you think
is right as the number of iteration is grown through the repetitve use
of an analysis command.
I hope that i have been clear enougth that this list will understand my point.
And i'm eager for your opinions on this toppic.
_________________________________________________________________
Votre correspondant a choisi Hotmail, un e-mail ultra sécurisé. Créez le votre
gratuitement !
http://www.windowslive.fr/hotmail/default.asp
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
