TalkChess.com

gladius wrote:A pretty amusing training game .

[pgn]
1. f3 d5 2. b3 c6 3. d3 Qd6 4. Bh6 Qxh6 5. Nh3 Qe6 6. Kf2 Qf5 7. Kg3 h5 8. Nf2
h4# {Black mates} 0-1
[/pgn]

I don't get why White didn't anticipate the simple h4# mate. Does it mean the training phase is only one ply deep?

I will be very curious to see how the MCTS + NN improve upon its tactical weakness.

I think If stockfish changes approach, it would only be to replace its evaluation with the NN while keeping the same search.

Those complaining about the match setup with the alphago have an obligation to become more productive and contribute CPU/GPU power

Thanks for doing this!

Daniel

Jhoravi wrote:

gladius wrote:A pretty amusing training game .

[pgn]
1. f3 d5 2. b3 c6 3. d3 Qd6 4. Bh6 Qxh6 5. Nh3 Qe6 6. Kf2 Qf5 7. Kg3 h5 8. Nf2
h4# {Black mates} 0-1
[/pgn]

I don't get why White didn't anticipate the simple h4# mate. Does it mean the training phase is only one ply deep?

No, but it's limited to 800 nodes. And if the network doesn't think the mate is a likely move, it might not be explored. I probably need to tune the UCT exploration parameter for chess vs go (it's using the original Leela Zero value right now).

cdani wrote:Nice! Congratulations!

I agree, Daniel, it's nice, but... how a mere mortal like me can help? And I'd eventually have my own learning engine, or my CPU time would help to build the learning stuff on remote?

Thanks to anybody who could clear my few and confused ideas...

I'm trying it. It's self-playing something but I can't understand the output. It's something like

1 <B D4> 2 <W Q4> 3 <B R17> 4 <W Q16>.........

and so on.

Well, I'll let it run to understand it while eating.

Interesting. But is it chess?

Rodolfo Leoni wrote:I'm trying it. It's self-playing something but I can't understand the output. It's something like

1 <B D4> 2 <W Q4> 3 <B R17> 4 <W Q16>.........

and so on.

Well, I'll let it run to understand it while eating.

Interesting. But is it chess?

LOL White won first game in 195 plies. No idea about the game anyway...
It'd be much more interesting if I could understand what happens...

Rodolfo Leoni wrote:

Rodolfo Leoni wrote:I'm trying it. It's self-playing something but I can't understand the output. It's something like

1 <B D4> 2 <W Q4> 3 <B R17> 4 <W Q16>.........

and so on.

Well, I'll let it run to understand it while eating.

Interesting. But is it chess?

LOL White won first game in 195 plies. No idea about the game anyway...
It'd be much more interesting if I could understand what happens...

I think you are running leela-zero, not lczero. Leela-zero is the Go player, and those look like Go moves .

lczero isn't quite ready for the distributed training portion yet, still working some bugs out of the system.

gladius wrote:

Rodolfo Leoni wrote:

Rodolfo Leoni wrote:I'm trying it. It's self-playing something but I can't understand the output. It's something like

1 <B D4> 2 <W Q4> 3 <B R17> 4 <W Q16>.........

and so on.

Well, I'll let it run to understand it while eating.

Interesting. But is it chess?

LOL White won first game in 195 plies. No idea about the game anyway...
It'd be much more interesting if I could understand what happens...

I think you are running leela-zero, not lczero. Leela-zero is the Go player, and those look like Go moves .

lczero isn't quite ready for the distributed training portion yet, still working some bugs out of the system.

Thanks Gary.
I suspected it was about Go because "moves" made no sense. I just found the "Leela Zero release" folder and I was very curious. Enough to keep me awake.

There'll be a meeting of Federal Regional Council on next saturday and I'll speak about LeelaCZero project. I'm sure it'll be very interesting for everybody.

I found another similar project
https://github.com/Zeta36/chess-alpha-zero

This happens a lot in the early (and sometimes later) training games in go as well. One player passes and doesn't realize the other one can win immediately by passing as well.

It will learn its way out of this.