Mah Jong help!

[Guest]

Archived from groups: rec.games.mahjong (More info?)

Dear all,

I am currently developping a computer player for Mah Jong, as the
dissertation of my Master of Informatics at Edinburgh University.
The program will be based on neural networks, but to achieve this I need
some data to train my network with!
When using neural network techniques, the data is a crucial factor to
get reasonable results (to get results at all actually).

As I am very new to Mah Jong myself, I am looking for experienced
players willing to share their precious expertise...

The task I am proposing would be like assessing the discards and melds
of a single player at some point in a deal, and name the most dangerous
and safest discards that one could make with respect to the data
available. An interface will be designed, so that the tiles will appear
on the screen, and the user would have to choose by clicking on the
tiles s/he feels are particulary safe/dangerous.

The interface will be developped in the next few days, in Java, so it
can be easily executed on any platform. Would anyone interested in
helping and having particular requirement regarding this program let me
know?


Thank you,

Regards,


Nico Toussaint

 

[Guest]

Archived from groups: rec.games.mahjong (More info?)

"Nicolas Toussaint" <s0343918@sms.ed.ac.uk> wrote

> I am currently developping a computer player for Mah Jong, as the
> dissertation of my Master of Informatics at Edinburgh University.
> The program will be based on neural networks, but to achieve this I need
> some data to train my network with!

You might find the strategy FAQ to be helpful in this regard...?

Tom

 

[Guest]

Archived from groups: rec.games.mahjong (More info?)

Nicolas Toussaint <s0343918@sms.ed.ac.uk> wrote in message news:<capke2$7pt$1@scotsman.ed.ac.uk>...
> I am currently developping a computer player for Mah Jong, as the
> dissertation of my Master of Informatics at Edinburgh University.
> The program will be based on neural networks, but to achieve this I need
> some data to train my network with!

Sounds terrific!

> The task I am proposing would be like assessing the discards and melds
> of a single player at some point in a deal, and name the most dangerous
> and safest discards that one could make with respect to the data
> available. An interface will be designed, so that the tiles will appear
> on the screen, and the user would have to choose by clicking on the
> tiles s/he feels are particulary safe/dangerous.

Just a piece of advice. There are DOZENS of Mahjong rules variations
which fall in three-four major groups. Strategies to play different
Rules differ, and, say, discard which is relatively "safe" in one case
may be "disaster" in other set of Rules. You'd better look for Rules
and Strategies in Tom Sloper's FAQ (you can find a link in other
thread).
So, BEFORE you start try to define which particular set of Rules you'd
be using for your program.

Best Regards,
Vitaly Novikov

P.S. Finishing with ONE set of Rules would count for your Master.
Finishing with ALL possible sets of Rules should count for PH.D. )

 

[Guest]

Archived from groups: rec.games.mahjong (More info?)

Tom Sloper wrote:
> You might find the strategy FAQ to be helpful in this regard...?

The idea behind neural network techniques is that they can
learn/generalise out of examples, without any other source of knowledge.
The knowledge must be 'extracted' out of experts through examples,
rather than dervied from rules.
I had a look to many FAQs and strategies, but unfortunately, they didn't
make me an experts in a few weeks time 😉

 

[Guest]

Archived from groups: rec.games.mahjong (More info?)

> The idea behind neural network techniques is that they can
> learn/generalise out of examples, without any other source of knowledge.
> The knowledge must be 'extracted' out of experts through examples,
> rather than dervied from rules.

You are missing the point. In the FAQ, you'll find examples of plays
made by an expert : http://www.sloperama.com/mahjongg/column.htm

By the way, you forgot to mention which set of rules you'll be using
for your experiment.

 

[Guest]

Archived from groups: rec.games.mahjong (More info?)

"Combo" <the_novikovs@hotmail.com> wrote in message
news:7c6538ba.0406162231.7f3f596c@posting.google.com...
> Nicolas Toussaint <s0343918@sms.ed.ac.uk> wrote in message
news:<capke2$7pt$1@scotsman.ed.ac.uk>...
> > I am currently developping a computer player for Mah Jong, as the
> > dissertation of my Master of Informatics at Edinburgh University.
> > The program will be based on neural networks, but to achieve this I need
> > some data to train my network with!
>
> Sounds terrific!
>
> > The task I am proposing would be like assessing the discards and melds
> > of a single player at some point in a deal, and name the most dangerous
> > and safest discards that one could make with respect to the data
> > available. An interface will be designed, so that the tiles will appear
> > on the screen, and the user would have to choose by clicking on the
> > tiles s/he feels are particulary safe/dangerous.
>
> Just a piece of advice. There are DOZENS of Mahjong rules variations
> which fall in three-four major groups. Strategies to play different
> Rules differ, and, say, discard which is relatively "safe" in one case
> may be "disaster" in other set of Rules. You'd better look for Rules
> and Strategies in Tom Sloper's FAQ (you can find a link in other
> thread).
> So, BEFORE you start try to define which particular set of Rules you'd
> be using for your program.

Hi Nicolas, to develop your own system (program, game, practice, etc.) of
mahjong, perhaps learning the basis of the game in its fullness is the most
important help you would need. So I agree with Vitaly, my advice to you is
to select a book of fully detailed rules of a variant, then start from
there!

>
> Best Regards,
> Vitaly Novikov
>
> P.S. Finishing with ONE set of Rules would count for your Master.
> Finishing with ALL possible sets of Rules should count for PH.D. )

I like that quote!

Cofa Tsui
www.iMahjong.com

 

[Guest]

Archived from groups: rec.games.mahjong (More info?)

> Tom Sloper wrote:
> > You might find the strategy FAQ to be helpful in this regard...?

"Nicolas Toussaint" <s0343918@sms.ed.ac.uk> wrote
> The idea behind neural network techniques is that they can
> learn/generalise out of examples, without any other source of knowledge.
> The knowledge must be 'extracted' out of experts through examples,
> rather than dervied from rules.
> I had a look to many FAQs and strategies, but unfortunately, they didn't
> make me an experts in a few weeks time 😉

All right, then. There are many examples in the strategy columns at
http://www.sloperama.com/mahjongg/column.htm. But which kind of mah-jongg is
your neural network going to be playing?

Your initial post said you wanted help with:

>>>assessing the discards and melds
>>>of a single player at some point in a deal, and name the most dangerous
>>>and safest discards that one could make with respect to the data
>>>available.

And that's stuff that is covered in the strategy FAQ, for a couple of
variants.
Anyway, good luck with your research. If I needed to find someone expert in
mah-jongg in Edinburgh, I'd start with Gareth Saunders.
http://www.garethjmsaunders.co.uk/news/contact.html

Cheers,
Tom

 

[Guest]

Archived from groups: rec.games.mahjong (More info?)

On Thu, 17 Jun 2004 19:04:39 +0100, Nicolas Toussaint
<s0343918@sms.ed.ac.uk> wrote:

>The idea behind neural network techniques is that they can
>learn/generalise out of examples, without any other source of knowledge.
>The knowledge must be 'extracted' out of experts through examples,
>rather than dervied from rules.

Excuse the possible naiveté - but ain't you creating something similar
to the in-built hint engine some Mahjong software uses?

Filipe

 

[Guest]

Archived from groups: rec.games.mahjong (More info?)

Filipe Silva wrote:
> Excuse the possible naiveté - but ain't you creating something similar
> to the in-built hint engine some Mahjong software uses?

A neural network would be quite different from the rules-based architecture that is (probably) used in most such games. It should be self-learning at a point. This would
make it a true AI.

Of course, if initially played by a novice, it might get worse! So, the idea (as I imagine it) would be to get it decently structured, then have it play against an expert
and emulate his choices. Later it could be made to learn from whoever (itself or the human) won each hand. At that point it could be viable as an opponent, and have the
potential to get quite good. In theory at least.
--
J. R. Fitch
Nine Dragons Software