Life In 19x19

You may have noticed that in the last weeks, the value of cryptocurrencies like Bitcoin has increased again.
This is good news for companies like NVidea and AMD who create the hardware for these miners.
And it is also good news for cloud service providers. It is not such a good news for ZBaduk though.

Our main cloud service provider has decided to raise its prices by 50%.
(Let's hope that Bitcoin soon crashes again, the only good use for a GPU is baduk mining.)

That does mean that the costs of running ZBaduk will also increase by 50% for now.
In the course of next week, ZBaduk will have to increase prices, as well.
(And we are considering to either limit or totally stop free use of the AI servers).

Q&A:

Why are we doing this ?
From its start ZBaduk has never made profit.
But we have to protect us against bigger losses.

What does it mean if you already have payed for a membership ?
Nothing will change for existing customers.
(You can see the expiry date of your membership in your "Account" page).

What does it mean for new customers ?
Prices will raise soon. If you already had plans to become a paying member,
you may want to do it before we increase prices.

Installed another update just a couple of minutes ago.

Again a lot of small changes:
Perhaps the most notable of all changes is the way ZBaduk displays color codes.

• From now on, the best move is shown in blue.
  The other moves are colored with different shades ranging from green to red. Green being the best move, red being the worst one.

In previous versions it was actually the number of playouts that was used to color the moves. For technical reasons, that is also how other tools do it (e.g. Lizzie). But it can be confusing. Anyway, from now on, ZBaduk uses the LCB. ZBaduk refers to this value as the "decision" value.

First of all, last week have been very heavy on our servers,
due to the promotions we did at the European Go Congress.
Next week we will introduce a new server which has more calculation power.

In mean time again a small release, with a little nifty feature:
Basically, it's an evaluation of the last played move.

Today's release puts a lot of effort in supporting KataGo as well as LeeLa Zero.
ZBaduk keeps a separate book keeping for each bot, but also merges the statistics to create an average value.
And we also show these 3 lines in the charts.

The chart will show up to 3 lines now. Red, Blue and Grey. Red and Blue are the 2 bots, and Grey is the average of them.
Because these 3 series are very close to each other, we decided to zoom the chart initially.
We did have to add a lot of normalization for LCB values, to keep the values between 0 and 100,
but anyway, when the LCB value has a low amount of visits, then we will use the winrate.

We did a tweak for LeeLa Zero which is pretty cool. When you play a move which is not proposed by LeeLa Zero,
and you move back up, and put it in "analyze this position" mode, it will actually force LeeLa Zero to consider your "creative" move.
And in return the statistics of that move will also be shown on the board.
If you play a really long variation, and that variation surpasses the main line in length.,
then after some time the variation will automatically be promoted and will become the main line,
'cause it makes sense in 99% of times.

The game list, now has an upload button, which also supports uploading multiple files.
There's also a fix for people who may have experienced login issues due to their e-mail address containing uppercase characters.
All of these issues should be fixed.

Problem with your averaging: both bots has H16 as top choice but B11 is chosen overall. H16 is shown in orange while B11 is in blue.

Hi Author, I have 2 questions.
1. I am not able to upload sgf and review full game, even though I bought a package with paypal. Is that an expected behaviour? Uploaded sgf doesn't contain any moves. So I have to relay move by move.
2. The number of playouts on the board circle is not the sum of LZ and KataGo playouts. Is that expected? For example, on the board I see 35k playouts for a move, but in the table below - LZ has 6k and KataGo has 7k playouts.

Cyan wrote:Problem with your averaging: both bots has H16 as top choice but B11 is chosen overall. H16 is shown in orange while B11 is in blue.

You are absolutely right. Something is wrong here.
It's hard to compare apples with oranges, but ZBaduk is trying to mix them to create multi-fruit juice.

What I am trying to say is, that Leela Zero and KataGo estimations are not easy to aggregate/merge/average. Some bots are more sensitive/optimistic than others when they generate a winrate. And sometimes only one bot rates a move, which makes it even more difficult to come up with an average.

One thing is for sure: We took a wrong turn when we tried to calculate a weighed average, that only makes things worse. Right now, ZBaduk uses the number of playouts, and often it does not really matter. But in examples like the one you posted, it screws things up.

A better approach is to benchmark the bots, and to find out how sensitive they are (a standard deviation) for moves they actually agree on. The result of this benchmark is just a simple standard deviation, which can be used to rescale future results. Once the results are all in the same range (i.e. same sensitivity) we can just take a simple average of them.
In this example there is a conversion factor of 1.1966 for converting KataGo scores to the range of leela zero. (which is the result of a quick benchmark).
It certainly is better than what we had before. So, that's probably what ZBaduk will use in its next release.

johnsmith wrote:Hi Author, I have 2 questions.
1. I am not able to upload sgf and review full game, even though I bought a package with paypal. Is that an expected behaviour? Uploaded sgf doesn't contain any moves. So I have to relay move by move.

Thank you for contacting me, it is certainly *not* expected behavior.
Could you send me the SGF file ? (here, or in a private message, or by e-mail: zbaduk361@gmail.com ).

Is it a handicap game file by any chance ?

Hi, thanks for answering. I don't remember the exact game, but it wasn't working for any game downloaded from gokifu, or just saved to sgf from Zen 7. It wasn't a handicap game.

I just installed another update of ZBaduk.
This release contains several small fixes for issues.
One of them is a rather important one.

(in)visible game records:

Apparently there was a bug that was hiding some of the game records of users.
In fact, you may have experienced issues uploading files to the server previously,
and it is possible that those files now suddenly appear.
For that matter, we also improved the responsibility of the game record list.

better merging of results

The 2nd major change, is how results of katago and leela zero are merged.
In short: it's getting better.
Longer version: it uses a conversion rate to scale the winrates of katago just a little.

ui settings

There is a configuration panel (it's in the "More..." dropdown list).
It contains some settings on how winrates/lcb values are displayed.

Some prefer winrates to be values in the range of [0-100] , but others actually prefer values in the range of -50 to +50.
And some prefer winrates to be expressed according to the active player, others prefer it to be from black's perspective.
And then some people prefer the rating chart to be zoomed (e.g. 40% to 60%), others prefer it to show the full range (e.g. 0 to 100)
All of this can now be configured.

johnsmith wrote:Hi, thanks for answering. I don't remember the exact game, but it wasn't working for any game downloaded from gokifu, or just saved to sgf from Zen 7. It wasn't a handicap game.

Perhaps the fix of "invisible games" which was installed today, may have an impact on the issue you described.

ZBaduk would appreciate some feedback about our services.
I created a short survey.
Your participation is certainly appreciated.

https://nl.surveymonkey.com/r/Z3HW26V

Services are great now as far as I am concerned, and thank you for that!
But there is one thing that I miss at the moment... There is a feature of KataGo that provides the difference in points between the two players, and I don't see that being implemented yet, or maybe I missed it somehow?

It's been a while, but just installed a new version.

There are several changes in this release.
Images say more than words:

That sums up most of it.