ZBaduk - LeeLa Zero and KataGo from your webbrowser

[xela]

For instance, there is a slight possibility that, with enough play-outs, Kata will start rating the 3-4 points better than the 4-4 points. I want to see if this is something that could happen with a million playouts of move 0.

Interesting question! I think it's unlikely but worth exploring. But I still think you could do it more efficiently.

A million playouts of move 0 gives you, what, 100k playouts of D4 (and similar for the other 4-4 points, but they should all get very similar evaluations), and a few tens of thousands of playouts of D3? You'd do better to just play D3 on the board, get 200k playouts of that position, do the same for D4, then see which has the higher evaluation. You get more (relevant) data for less computation and less of your time.

The problem is that if D3 has a lower policy value, then it could take some massive number of total playouts before D3 starts getting a significant proportion of that total. That's why, if you're interested in a specific move, it's usually better to play that move on the board rather than analyse from the previous position.

[zermelo]

My best and most honest answer is that I'm incurably neurotic.

My slightly less honest but a bit more practical answer is that IMO, the easier it is to get to a million playouts (for any move, not just move 0), the better. For instance, there is a slight possibility that, with enough play-outs, Kata will start rating the 3-4 points better than the 4-4 points. I want to see if this is something that could happen with a million playouts of move 0.

I think 1 million playouts costs something in the order of 1 USD computed on the cloud computers. I find a bit unreasonable to expect to get unlimited background computation time with Zbaduk's prices. I think the system is meant for interactive game reviews. You can always set up a cloud-computing system of your own. I've for instance used this and modified it at some points https://github.com/bsteuber/lizzie-gcloud-setup.

[xela]

For instance, there is a slight possibility that, with enough play-outs, Kata will start rating the 3-4 points better than the 4-4 points. I want to see if this is something that could happen with a million playouts of move 0.

OK, I can answer that one for you. The miracle didn't happen, at least not today. Maybe it needs ten million playouts? :-)

[Maharani]

I think 1 million playouts costs something in the order of 1 USD computed on the cloud computers.

I had no idea, and now feel bad for asking...

EDIT: I've managed to improve my internet connection. Sixty minutes is enough for my purposes by now. On the other hand, I still wouldn't mind joining, say, a higher pay-level sponsorship tier without automatic disconnection for people who need high playouts... Just a suggestion

[spook]

I still wouldn't mind joining, say, a higher pay-level sponsorship tier without automatic disconnection for people who need high playouts... Just a suggestion

Setting a correct price is difficult. Some chinese vendors use credit systems, and rent services per minute.
When ZBaduk started out that was impossible, because it had to provision servers in advance.
It would be complicated to introduce it now.

I can only acknowledge that prices are indeed a bit too low.
(some chinese vendors charge 60$/mo, and e.g. an AWS GPU instance costs about 0.8$/hr)
At the same time, I guess that may be what makes ZBaduk stand out.
The very fact that it is cheaper to use ZBaduk than to buy your own hardware.

January was actually a good month. One of the first months where:
membership_sales > hardware_costs.

So, ZBaduk is happy !
Crunch those numbers !

[y501]

Hi,

¿Can you get verified to receive contributions from Brave users? I just gave 5$ worth of BAT to OGS and I'd like to give the BAT I get monthly to you.

Also, do you still need someone to translate your site to Korean?

[Maharani]

January was actually a good month. One of the first months where:
membership_sales > hardware_costs.

That's great to hear - congratulations!

So, ZBaduk is happy !
Crunch those numbers !

You're the best. Thank you once again for the service you offer. <3

EDIT: Maybe I over-interpreted "crunch those numbers"? When I'm "just idle", I still time out automatically after at least sixty and at most seventy minutes.

[spook]

EDIT: Maybe I over-interpreted "crunch those numbers"? When I'm "just idle", I still time out automatically after at least sixty and at most seventy minutes.

@Maharani , oopz, I only just realize now, that may have been confusing. Sorry about that.
What I did mean: "don't feel guilty". ; What I didn't mean: "all limitations are gone".

I will look further into the ownership-toggle thing which you mentioned though. (in a couple of weeks).
In mean time I hope to add some other features which you may like.

Some transparency about what can be expected in the coming weeks:

Right now, there are a couple of technical bottlenecks in the architecture of ZBaduk, which I am trying to enhance in the next release. Even though they don't add new functionality, it will be important for future development.

There will be a "game properties panel" soon, which will allow to edit game results, dates, rules and other properties. So, that will be coming in about 2 weeks I think.

Part of the plan behind the game properties panel, is to make the step towards "sharing" of reviews. e.g. you played a game with somebody, you review it, and you want to send him a link to the review by mail. I don't know exactly when this will be ready, but perhaps in about a month.

I created an experimental voting page for upcoming features:
http://www.tricider.com/brainstorming/3Z1ljaALAud
Feel free to add features or leave votes. Feel free to vote on multiple ideas.

[Maharani]

EDIT: Maybe I over-interpreted "crunch those numbers"? When I'm "just idle", I still time out automatically after at least sixty and at most seventy minutes.

@Maharani , oopz, I only just realize now, that may have been confusing. Sorry about that.
What I did mean: "don't feel guilty". ; What I didn't mean: "all limitations are gone".

I will look further into the ownership-toggle thing which you mentioned though. (in a couple of weeks).

Got it, thanks for clarifying.

I hope you don't mind me re-asking a couple of questions you may have missed:

If this issue was resolved, would using the ownership button signal to the server that I'm just idle rather than completely idle? I assume it would depend on whether workaround 1 or 2 was implemented?

[...]

On an unrelated note, ZBaduk is the only website I've visited in a very long time for which it is not sufficient to type zbaduk.com into my browser's address line (regardless of OS or browser), which produces a completely blank page (it doesn't seem to be a 404 type of error page, though). I need to specifically type https://zbaduk.com to reach the website when my browser cache is clear. Only a minor inconvenience, but a pretty non-standard one.

[...

Up to] move 40, the calculations are saved to the server regardless [... .] This does not seem to be the case for calculations based on komi 7. I don't really care, but I wonder if this a fluke or intentional?

[Gomoto]

I would appreciate settings to turn of the winrate and only show the katago score very much.

I use score only on my pc because I like it, but at the club while using a tablet it would be also an advantage because of the smaller screen size.

[spook]

... ZBaduk is the only website I've visited in a very long time for which it is not sufficient to type zbaduk.com into my browser's address line (regardless of OS or browser), which produces a completely blank page

Thank you for informing me about this --> should be fixed !

This does not seem to be the case for calculations based on komi 7. I don't really care, but I wonder if this a fluke or intentional?

Of course in the future, it has to be rolled out to other rulesets and komi values, but right now that cache only supports 7.5+chinese rules.

What's holding me back? --> The right thing to do: a redesign of the caching system, to migrate it to a real database.
Why ? --> because it uses too much server memory and therefor should run on its own dedicated server. (takes about 700MB RAM/cache)

If this issue was resolved, would using the ownership button signal to the server that I'm just idle rather than completely idle?

As you stated, it depends on the solution. Right now, clicking show/hide triggers communication with the back-end. As it looks now, it is likely that this won't be the case in the future. It will just show/hide the statistics, and won't communicate anything to the server anymore. So, it will have no more effect on "idleness". (That's good news, right? ... but only if everything goes according to plan.)

[Maharani]

So, it will have no more effect on "idleness". (That's good news, right? [...])

Not for me, since the main reason I suggested this "bug fix" is that I intended to use the toggle to communicate "just-idleness" to the server in order to cheat the sixty-minute calculation limit. But it's all good.

[spook]

I would appreciate settings to turn of the winrate and only show the katago score very much.
I use score only on my pc because I like it, but at the club while using a tablet it would be also an advantage because of the smaller screen size.

I agree, ZBaduk urgently needs an option to show/hide columns in its tables.

Given the fact that you only use score,
perhaps the following could be even more valuable for you ?

score on goban.jpg (162.17 KiB) Viewed 11882 times

It could be in the next front-end release,
I just need to figure out how to switch between modes.
Perhaps a button to toggle between the score/winrate views.

button to toggle it.jpg (31.32 KiB) Viewed 11882 times

The buttons could look like this :

buttons to add or remove statistics.jpg (3.59 KiB) Viewed 11880 times

Probably in a week or 2 ... (and of course only for KataGo).

[Maharani]

Yet another question... sorry. How come the winrates displayed on the board for 7 komi are different (lower) than the winrates given in the chart? I don't remember this happening for 7.5 komi reviews.

EDIT: I think the reason is that the displayed-on-the-board values are an average of KataGo 7 komi New Zealand rules and Leela Zero 7.5 komi Chinese rules, except that the Leela Zero values are hidden in the chart?

[y501]

I would appreciate settings to turn of the winrate and only show the katago score very much.
I use score only on my pc because I like it, but at the club while using a tablet it would be also an advantage because of the smaller screen size.

I agree, ZBaduk urgently needs an option to show/hide columns in its tables.

Given the fact that you only use score,
perhaps the following could be even more valuable for you ?

score on goban.jpg

It could be in the next front-end release,
I just need to figure out how to switch between modes.
Perhaps a button to toggle between the score/winrate views.

button to toggle it.jpg

The buttons could look like this :

buttons to add or remove statistics.jpg

Probably in a week or 2 ... (and of course only for KataGo).

That looks amazing