Twinbot-0.3 is a patched GNU Go 3.7.8
This is the first version, it has some known weaknesses.
The development version is Twinbot
It runs on kgs under the name Twinbis.
Instead of doing just a move generation for one color, it generates moves for both colors (hence the name) then apply some nice heuristic to find the best move.
Let's be more precise:
- If one opponent top move is near one of our top move, add a bonus, then play the best move :)
On a 19x19, this changes roughly 20% of the generated moves.
Why? Sometimes, when playing against itself, GNU Go (let's say Black) plays terrible tenuki, and gets in trouble as soon as GNU Go (White) plays.
If they could speak, in a teaching game, white would have given the information to black:
- White: Eh, you cannot tenuki, i cut/trap/crush you here.
- Black: Lol, that's true, i have not seen that!
Things are somewhat different in the engine, but the idea is here, pass information from opponent point of vue, and try to improve the game evaluation.
How? hmm, it is my very first C hack, so please be tolerant. http://trac.gnugo.org/gnugo/ticket/104 contains the first draft and regressions test results. Since then, the semeai pb has been nicely fixed (see Twinbot).
Is it really good? I hope so, at the cost of 50% slow down. It is a work in progress, and there is room for lots of improvements, code reordering, clean-up... It appears to be rather good for playing more big moves in end chuban / early yose on 19x19 board.
On a 9x9 board, preliminary results seems good against Lazarus, and Viking 5, compared with "standard" GNU Go.