toss-devel

[Toss-devel] TOSS NLP

[Patrick H. <pat...@ho...>]

Dear TOSS devs!

I'm impressed especially by the ideas related to the AI of this project, especially how it learns to play arbitrary board games. I'm interested in developing some sort of NLP interface (or some sort of NLP board game at the beginning ) for TOSS based on the learning techniques this AI uses for board games and contribute it to this project.
I think that the learning technique TOSS uses could also be a breakthrough in natural language processing of simple sentences.

I managed to compile TOSS on Tiny Core Linux, but how can I compile changed toss example files in such a way that the browser interface will be adapted to the new file, and the TOSS AI will have created tactics for my modified toss file
such that I can test the AI. Is it described in the reference pdf? I will read it the next days.

Another question: Is it possible to hide game knowledge from the AI? If not, no big deal,
since I can model unknown knowledge by the "movements" of other players.

And can you add me to the mailing list please?

Best regards,
Patrick Hammer
 		 	   		  

Re: [Toss-devel] TOSS NLP

[Lukasz S. <luk...@gm...>]

Hello Patrick,

Thank you for your enthusiasm! You may have a look at our subsystem for
richer language processing:

http://toss.sourceforge.net/term.html

Let us know if you like it. If you do, I will likely document it, and write
a tutorial. If you start learning it, we can discuss the details on the
mailing list.

The learning aspect of game playing is not that well developed yet. We had
plans to implement weight learning for "board" features but haven't got
time to work on it. I leave the webclient and Toss server issues to the
other email.

Regards.

Re: [Toss-devel] TOSS NLP

[Lukasz K. <luk...@gm...>]

Hi Patrick.

Let's start with the technical problems.

> I managed to compile TOSS on Tiny Core Linux, but how can I compile changed
> toss example files in such a way that the browser interface will be adapted
> to the new file, and the TOSS AI will have created tactics for my modified
> toss file such that I can test the AI. Is it described in the reference pdf?

The first thing I do when creating a new .toss file is like this
  ./TossServer -f examples/Connect4.toss -play 1
where of course you replace "examples/Connect4.toss" with your file.
This starts a computer-vs-computer self-play with 1s per move limit.

There is no bash-interface for human-vs-computer playing. It would be
very easy to do (we might get to it if you insist), but it'd probably
be quite annoying: ascii display only goes so far, and typing an exact
move rule each time gets annoying very fast. That's why we focused on
a web interface.

To start the web interface, you must have both TossServer and
Client/JsHandler.js compiled - the standard "make" compiles them both,
but make sure they're there. Then you just start the server
  ./TossServer -p 1234
and go in your browser to localhost:1234. That should (by default)
show a local copy of tplay.

Now - to get your game there, you must do a few things. The first
thing to remember is that we designed the web stuff to work offline
too - so all games are pre-compiled to javascript. In the Makefile,
this is done by the rules that put TOSSEXFILES (examples/*.toss) into
Resources ml, which then get compiled into javascript. No need to
change anything there, just put your .toss file under examples/ and do
"make clean; make" to re-compile. Then, you must go inside
Client/JsHandler ml and look at the gSel_predef_games record - that's
how game names are associated with toss files. So just add a line
there for your file and your game name. And that's it for the basics:
if you re-start the server and open localhost:1234/?game=YourGame, you
should get a basic UI for your game. Now - making this UI really
usable might require a bit more work, but I hope this can at least
help you get started :). (If you see strange JS bugs in your browser,
it might be because you are not using the right version of
ocaml-to-javascript compiler. OPAM should give you the right one, but
if you see problems, just let us know.)

> I'm impressed especially by the ideas related to the AI of this project,
> especially how it learns to play arbitrary board games. I'm interested in
> developing some sort of NLP interface (or some sort of NLP board game at the
> beginning ) for TOSS based on the learning techniques this AI uses for board
> games and contribute it to this project.
> I think that the learning technique TOSS uses could also be a breakthrough
> in natural language processing of simple sentences.

That sounds very interesting :). We tried a bit with our own NLP, as
the other Lukasz already wrote you, it is all in Term. If I were to do
that again now, I think I'd not try to do parsing myself but use an
existing parser (probably zpar, see zpar on sourceforge), and work
forward from the resulting constituency trees. But that's just my
opinion, maybe the other Lukasz does not agree ;).

> Another question: Is it possible to hide game knowledge from the AI? If not,
> no big deal, since I can model unknown knowledge by the "movements" of other players.

Do you mean hiding from the player in general, or just making the AI
weaker? To make the AI stupid, you can just give it little time, it
moves at random then. Full support for games with imperfect
information was high on our list for some time, but I'm afraid it
never got done.

My best and keep in touch if you need any help!

Lukasz Kaiser

Re: [Toss-devel] TOSS NLP

[Lukasz S. <luk...@gm...>]

On Mon, Mar 10, 2014 at 1:23 AM, Lukasz Kaiser <luk...@gm...>wrote:

>
> That sounds very interesting :). We tried a bit with our own NLP, as
> the other Lukasz already wrote you, it is all in Term. If I were to do
> that again now, I think I'd not try to do parsing myself but use an
> existing parser (probably zpar, see zpar on sourceforge), and work
> forward from the resulting constituency trees. But that's just my
> opinion, maybe the other Lukasz does not agree ;).
>

I totally agree, as an approach to consuming unrestricted natural language.
I held that opinion even before starting the work underlying
http://toss.sourceforge.net/term.html (to tie loose ends of ideas we
explored before). However, if you are thinking about doing a controlled
English style game, with the vocabulary and grammar built from the ground
up, it might still be a good framework and parser to try. If "Term" was
better integrated with Toss as a whole, it would certainly be appealing in
that spirit. But especially as it's not so tightly integrated, I do support
starting from scratch on a subsystem that interfaces with the output of a
broad-coverage grammar/parser.

Cheers.