Life In 19x19

Because she deserves a thread

https://www.reddit.com/r/baduk/comments ... ero_major/

This is basically aimed towards analysis.

And you are so right, it deserves a lot of hype and improvements at the same time, as can be read in the Reddit thread. I was thinking maybe the author wants to create an announcement in the annountsment forum here.
Analysis will be joy! Mouseover and UI need some further thinking, but it is a good beginning. Awesome times.

At times the UI is so hyper it feels like I'm being brainwashed with subliminal messages.

Lizzie is such a great go learning tool. Watch for yourself:

(I am not a streamer by the way, I just wanted to show the amazing team LeelaZ & Lizzie.
So don't bother my bad english and my bad comments )

Amazing!! I love it.

That's fantastic! Sadly, my GPU isn't really good enough for these new bots. I'm tempted to buy a nice graphics card just for playing with this, but the cryptocurrency craze has kept prices ridiculously high.

Lizzie Version 0.4 released

Downloads here

https://raw.githubusercontent.com/featu ... screen.png

Changelog:
- Winrate graph
- Winrate indicator
- Variation tree
- Pondering indicator
- Lizzie will now give you a proper error message if you are not using the correct Leela Zero binary, instead of silently failing
- Fixed Windows-related issues that would not let Lizzie start
- New game creation / handicap game support
- Use system default look and feel for dialog boxes, instead of the Java Swing
- Next move indicator

jeromie wrote:That's fantastic! Sadly, my GPU isn't really good enough for these new bots. I'm tempted to buy a nice graphics card just for playing with this, but the cryptocurrency craze has kept prices ridiculously high.

Maybe time to upgrade my Pentium 4 desktop! On my work laptop with latest 15 block network I get about 10 playouts a second, a long way off mr baduk1. But even with such few playouts it's still pretty strong, so long as ladders aren't involved. e.g. I couldn't really reproduce this ladder-dependence goodness of the small avalanche from AG, though latest 15 block was better than an older 10 block, and when analysing my taisha from last PGETC game LZ wanted to keep going down a ladder disaster road for a while.

Is it possible to have it "batch" analyze a game? What I mean is to give it a game, let the system analyze it, then view the results later.

hyperpape wrote:Is it possible to have it "batch" analyze a game? What I mean is to give it a game, let the system analyze it, then view the results later.

No, but Go Review Partner is already good for that.

Lizzie 0.4! https://github.com/featurecat/lizzie/releases/tag/0.4

Even though my computer is crap (<10 visits per second on 15 block network), Leela Zero (network #126) plus Lizzie still makes it a pleasure to review games quickly and find interesting things. As an example, I looked at AlphaGo vs Lee Sedol game 2, here's a few highlights.

The famous move 37 shoulder hit, apparently only 1 in 10,000 or whatever chance of human to play here based on the policy network trained on human games. But it's Leela Zero's #3 choice and found with fewer than a dozen visits. "Obvious, duh!" to these strong Zero bots . For comparison Leela 0.11 didn't even consider it (and considered 36 other moves) after 360k simulations. The other moves it likes are also ones that later happened in the game: the shoulder hit lower side, preparing connection on left, and the nozomi (more examples from AG). [black or white circle indicates next move in real game, red circle indicates LZ's best move, with strongest green down to red for bad ones, top number is win%, bottom number is # visits]. After a few hundred more visits it's still liked but not #1, after 100k visits who knows?: With the lower side fighting after a while LZ finds the crude but effective atari that Lee Sedol should have played (as found by Shi Yue iirc); this can involve a ladder or a net which she just about seems to understand after a few thousand visits. Or with that remarkable indirect response to Lee's top side invasions which elicited praise from Myungwan Kim as revealing great strength, the instinct is a local move like cap or defend corner but it finds it within 100 visits and thinks it's clearly best after ~1000.

I'm really glad you are enjoying using Lizzie for analysis! Analyzing previously played games is my main use case so I worked on a bunch of features for 0.4 that try to make that a more pleasant experience. It is really addictive now and I find that when I have trouble summoning the energy to play a game, "I'll be able to analyze it with Lizzie afterwards" often provides me with sufficient motivation.

I should also add that the principal author featurecat, in addition to writing the app in the first place, has been really great about accepting other people's work. Like Leela Zero, it's a great advertisement for collaborative open source software.

The famous move 37 shoulder hit...

It seems that this move is less likely to be played after more playouts (Lizzie 0.4 with 3f6c8)
here
and here

I think this is probably my favourite LeelaZero utility so far.