[Computer-go] Beating old ManyFaces at 29 handicap stones

["Ingo Althöfer"]

Hello Darren,

very interesting thoughts and proposals.

> This is an interesting proposal. Especially if the same bot can play
> strongly at all handicaps. (I.e. winning at 29 handicap stones, but
> losing at 20 handicap stones, tells us the programmers have just written
> a good 29-stone-handicap fuseki.)
> 
> One suggestion, on that theme, is to require programs to *qualify* by
> beating MF at increasing handicaps, on KGS or somewhere. With the
> condition it has to be the same binary and database at each handicap. :-)

That is a good idea. In concrete, one might ask applicants to win
(against mfgo1998) at 17, 19, 21, 23, 25, 27, 29.

Even if no bot would get the main prize for handicap 29 until December 2020, 
I might give a smaller amount of money to the maker of that bot that achieves
at least three wins out of 5 games at the highest level from the list above.
 
> ... It seems to me like an
> MCTS program (with some suitable dynamic komi cleverness) could do
> rather well in this kind of challenge.

My expectations are: 
(i) the full prize will be won clearly before end of 2020.
(ii) the winning bot will be based on Monte Carlo TS.
(iii) the winning bot will also use some sorts ofopponent modelling.
(iv) the scene will learn from the process also for handicap play against
     other opponents (including humans).

> In fact I think a bot is more likely to beat MF at 29 stones than beat
> Martin Mueller in an even game :-)

Uhh. 

> ...hhmmm, that might make an interesting poll to setup!

I think that Martin - when he plays - will be beaten by a top bot within
the next two years, even at slow time controls.

> I guess your motivation is to push some progress in dynamic komi (and
> similar) research?

In dynamic komi, and at least to the same amount in opponent modelling.
 
Ingo.
-- 
Empfehlen Sie GMX DSL Ihren Freunden und Bekannten und wir
belohnen Sie mit bis zu 50,- Euro! https://freundschaftswerbung.gmx.de