Re: [Computer-go] LZ-1a198f-t1-v2600 on CGOS

[Andy]

With the recent upgrade to 10 block 128 filter net for LZ a few more people
have started running LZ bots on CGOS. I'm running the last 6x128 net as
LZ-b3b00c-t1-v3200, and it is using a recent build from the next branch
(v0.12-24-g4a77da5). I plan to make this an anchor that runs a long time. I
also tried running a full thinking time bot as LZ-1a198f-t2-nolim, but it
caused my computer to lockup and lose on time. I'll try debugging that
later, and in the meantime avoid nolimit bots.

Some other LZ bots active now:
LZ-5b9-t4, LeelaZero-10B-w5b9, LeelaZ_1080_ae205d. I don't know anything
about these but they appear to be recent 10x128 nets, and they use a good
amount of time to play so probably don't have a playout/visit cap. I guess
I'll wait awhile to see if more LZ bots appear before running any new
10x128 nets myself.

- Andy aka KillerDucky




2018-03-04 22:38 GMT-06:00 Andy :

> See below Cross-post from https://www.reddit.com/r/cbaduk/comments/823ekq/
> lz1a198ft1v2600_on_cgos
>
> I'm volunteering to run some LZ bots on CGOS (the main person running them
> before has decided to stop). The first one, LZ-1a198f-t1-v2600, will get
> 100 games soon.
>
> I decided to use -v2600 because it uses about the same amount of GPU time
> as -p1600 but it is stronger. Full command line: ./leelaz -v 2600
> --noponder -r 1 -g -t 1 -d -w 1a198f -l LZ-1a198f-t1-v2600.log
>
> I plan to eventually run two bots at a time. One as an anchor that runs
> for a long time, and another as the most recent net just up to 100 games to
> get a BayesElo. This might mean skipping some nets but I don't think it's
> necessary to get every single one. Some other bot authors have complained
> about too many LZ clones on CGOS.
>
> I also plan to occasionally run a net with full thinking time (remove -v2600
> --noponder, and use -t2 which seems close to optimal on my machine). I
> have a GTX960, when I run this test I'll post some information on how many
> playouts it gets in the games.
>
> - Andy aka KillerDucky
>
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] LZ-1a198f-t1-v2600 on CGOS

[Andy]

See below Cross-post from
https://www.reddit.com/r/cbaduk/comments/823ekq/lz1a198ft1v2600_on_cgos

I'm volunteering to run some LZ bots on CGOS (the main person running them
before has decided to stop). The first one, LZ-1a198f-t1-v2600, will get
100 games soon.

I decided to use -v2600 because it uses about the same amount of GPU time
as -p1600 but it is stronger. Full command line: ./leelaz -v 2600
--noponder -r 1 -g -t 1 -d -w 1a198f -l LZ-1a198f-t1-v2600.log

I plan to eventually run two bots at a time. One as an anchor that runs for
a long time, and another as the most recent net just up to 100 games to get
a BayesElo. This might mean skipping some nets but I don't think it's
necessary to get every single one. Some other bot authors have complained
about too many LZ clones on CGOS.

I also plan to occasionally run a net with full thinking time (remove -v2600
--noponder, and use -t2 which seems close to optimal on my machine). I have
a GTX960, when I run this test I'll post some information on how many
playouts it gets in the games.

- Andy aka KillerDucky
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Crazy Stone is back

[Andy]

 Rémi, can you share any details about how you are training your network?
Are you doing self-play games? Do you have access to a large number of GPUs?



2018-02-28 13:04 GMT-06:00 David Wu :

> It's not even just liberties and semeai, it's also eyes. Consider for
> example a large dragon that has miai for 2 eyes in distant locations, and
> the opponent then takes one of them - you'd like the policy net to now
> suggest the other eye-making move far away. And you'd also like the value
> net to distinguish the three situations where the whole group has 2 eyes
> even when they are distant versus the ones where it doesn't.
>
> I've been doing experiments with somewhat smaller neural nets (roughly 4-7
> residual blocks = 8-14 layers), without sticking to an idealized "zero"
> approach. I've only experimented with policy nets so far, but presumably
> much of this should also transfer to a value net's understanding too.
>
> 1. One thing I tried was chain pooling, which was neat, but ultimately
> didn't seem promising:
> https://github.com/lightvector/GoNN#chain-pooling
> It solves all of these problems when the strings are solidly connected. It
> helps also when the strings are long but not quite solidly connected too,
> the information still propagates faster than without it. But of course, if
> there are lots of little strings forming a group, diagonal connections,
> bamboo joints, etc, then of course it won't help. And also chain pooling is
> computationally costly, at least in Tensorflow, and it might have negative
> effects on the rest of the neural net that I don't understand.
>
> 2. A new thing I've been trying recently that actually does seem
> moderately promising is dilated convolutions, although I'm still early in
> testing. They also help increase the speed of information propagation, and
> don't require solidly connected strings, and also are reasonably cheap.
>
> In particular: my residual blocks have 192 channels, so I tried taking
> several of the later residual blocks in the neural net and making 64 of the
> channels of the first convolution in each block use dilated convolutions
> (leaving 128 channels of regular convolutions), with dilation factors of 2
> or 3. Intuitively, the idea is that earlier blocks could learn to compute
> 2x2 or 3x3 connectivity patterns, and then the dilated convolutions in
> later residual blocks will be able to use that to propagate information
> several spaces at a time across connected groups or dragons.
>
> So far, indications are that this works. When I looked at it in various
> board positions, it helped in a variety of capturing race and
> large-dragon-two-eye-miai situations, correctly suggesting moves that the
> net without dilated convolutions would fail to find due to the move being
> too far away. Also dilated convolutions seem pretty cheap - it only
> slightly increases the computational cost of the net.
>
> So far, I've found that it doesn't significantly improve the overall loss
> function, presumably because now there are 128 channels instead of 192
> channels of ordinary convolutions, so in return for being better at
> long-distance interactions, the neural net has gotten worse at some local
> tactics. But it also hasn't gotten worse the way it would if I simply
> dropped the number of channels from 192 to 128 without adding any new
> channels, so the dilated convolutions are being "used" for real work.
>
> I'd be curious to hear if anyone else has tried dilated convolutions and
> what results they got. If there's anything at all to do other than just add
> more layers, I think they're the most promising thing I know of.
>
>
> On Wed, Feb 28, 2018 at 12:34 PM, Rémi Coulom  wrote:
>
>> 192 and 256 are the numbers of channels. They are fully connected, so the
>> number of 3x3 filters is 192^2, and 256^2.
>>
>> Having liberty counts and string size as input helps, but it solves only
>> a small part of the problem. You can't read a semeai from just the
>> liberty-count information.
>>
>> I tried to be clever and find ways to propagate information along strings
>> in the network. But all the techniques I tried make the network much
>> slower. Adding more layers is simple and works.
>>
>> Rémi
>>
>> - Mail original -
>> De: "Darren Cook" 
>> À: computer-go@computer-go.org
>> Envoyé: Mercredi 28 Février 2018 16:43:10
>> Objet: Re: [Computer-go] Crazy Stone is back
>>
>> > Weights_31_3200 is 20 layers of 192, 3200 board evaluations per move
>> > (no random playout). But it still has difficulties with very long
>> > strings. My next network will be 40 layers of 256, like Master.
>>
>> "long strings" here means solidly connected stones?
>>
>> The 192 vs. 256 is the number of 3x3 convolution filters?
>>
>> Has anyone been doing experiments with, say, 5x5 filters (and fewer
>> layers), and/or putting more raw information in (e.g. liberty counts -
>> which makes the long string problem go away, if I've understood
>> correctly what that is)?
>>
>> Darren
>

[Computer-go] L011 bots on CGOS are configured wrong

[Andy]

Hoping to contact whoever is running these bots.

https://www.reddit.com/r/cbaduk/comments/7nacja/l011_bots_on_cgos_are_configured_wrong/

Copy/paste of my post on reddit:

Who is running the L011 bots on CGOS? I'm assuming they are Leela v0.11.
They are not configured to remove all dead stones from the board. I don't
know the command line options for Leela, does anyone know how to get it to
do that? Also when you fix it please make new accounts because it's going
to mess up all the ratings otherwise.

http://www.yss-aya.com/cgos/viewer.cgi?19x19/SGF/2017/12/31/342411.sgf

Also can you run a test to see how much time it takes to do X playouts in
Leela vs in LeelaZero? I don't think those numbers are directly comparable.


And roy7's helpful addition:

What the bot operator has to do is change the cgos wrapper program to
always send the client kgs-genmove_cleanup instead of "genmove". I think
it's a line in gtpengine.py that looks like this:

result = self._sendListResponseCommand("kgs-genmove_cleanup " + gtpColour)

Yes, people running Leela without this will lose games they should have won
because otherwise Leela is playing Chinese but the server is playing
Tromp-Taylor. TT and Chinese are similar enough just forcing
kgs-genmove_cleanup will work around it.
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] AGZ Policy Head

[Andy]

Is there some particular reason AGZ uses two 1x1 filters for the policy
head instead of one?

They could also have allowed more, but I guess that would be expensive? I
calculate that the fully connected layer has 2*361*362 weights, where 2 is
the number of filters.

By comparison the value head has only a single 1x1 filter, but it goes to a
hidden layer of 256. That gives 1*361*256 weights. Why not use two 1x1
filters here? Maybe since the final output is only a single scalar it's not
needed?
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] mcts and tactics

[Andy]

How do you interpret this quote from the AGZ paper?
"Surprisingly, shicho (“ladder” capture sequences that may span the whole
board) – one of the first elements of Go knowledge learned by humans – were
only understood by AlphaGo Zero much later in training."

To me "understood" means the neural network itself can read at least some
simple whole board ladders, ladder breakers, and ladder makers. I would
find it a large oversell if they just mean the MCTS search reads the ladder
across the whole board.



2017-12-19 18:16 GMT-06:00 Stephan K :

> 2017-12-20 0:26 UTC+01:00, Dan :
> > Hello all,
> >
> > It is known that MCTS's week point is tactics. How is AlphaZero able to
> > resolve Go tactics such as ladders efficiently? If I recall correctly
> many
> > people were asking the same question during the Lee Sedo match -- and it
> > seemed it didn't have any problem with ladders and such.
>
> Note that the input to the neural networks in the version that played
> against Lee Sedol had a lot of handcrafted features, including
> information about ladders. See "extended data table 2", page 11 of the
> Nature article. You can imagine that as watching the go board through
> goggles that put a flag on each intersection that would result in a
> successful ladder capture, and another flag on each intersection that
> would result in a successful ladder escape.
>
> (It also means that you only need to read one move ahead to see
> whether a move is a successful ladder breaker or not.)
>
> Of course, your question still stands for the Zero versions.
>
> Here is the table :
>
> Feature # of planes Description
>
> Stone colour3   Player stone /
> opponent stone / empty
> Ones1   A constant plane
> filled with 1
> Turns since 8   How many turns
> since a move was played
> Liberties   8   Number of
> liberties (empty adjacent points)
> Capture size8   How many opponent
> stones would be captured
> Self-atari size 8   How many of own
> stones would be captured
> Liberties after move8   Number of
> liberties after this move is played
> Ladder capture  1   Whether a move at this
> point is a successful ladder capture
> Ladder escape   1   Whether a move at
> this point is a successful ladder escape
> Sensibleness1   Whether a move is
> legal and does not fill its own eyes
> Zeros   1   A constant plane
> filled with 0
>
> Player color1   Whether current
> player is black
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm

[Andy]

Google has already announced their next step -- Starcraft2. But so far the
results they published aren't mind blowing like these.


2017-12-19 9:15 GMT-06:00 Fidel Santiago :

> Hello,
>
> I was thinking about this development and what it may mean from the point
> of view of a more general AI. I daresay the next experiment would be to
> have just one neural net playing the three games, right? To my
> understanding we still have three instances of the same *methodology* but
> not yet a single one playing different games.
>
> Best regards,
>
> Fidel Santiago.
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] kgs-chat command not working

[Andy]

I'm trying to implement the kgs-chat command but I'm getting a crash in
kgsGtp-3.5.22. Everything works fine and I can chat with the bot. But when
a match starts that's when it crashes. Here is my gtp log. Any ideas? Or
can someone paste me an example log that works or point me to example code?

name
= myname

version
= myversion

kgs-chat private yoyoma .
= hello

boardsize 19
=

kgs-time_settings none
=

boardsize 19
java.lang.NullPointerException
at com.gokgs.client.gtp.GtpConvo.handleChat(GtpConvo.java:70)
at com.gokgs.client.gtp.GtpConvo.handleConvoEvent(GtpConvo.java:64)
at org.igoweb.util.Multicaster.handleEvent(Multicaster.java:107)
at org.igoweb.util.Emitter.emit(Emitter.java:85)
at org.igoweb.igoweb.client.CChannel.appendChat(CChannel.java:249)
at com.gokgs.client.KCGame.extractChats(KCGame.java:393)
at com.gokgs.client.KCGame.applyTreeEvent(KCGame.java:376)
at com.gokgs.client.KCGame.readGameUpdate(KCGame.java:366)
at org.igoweb.igoweb.client.CGame.handleMessage(CGame.java:163)
at com.gokgs.client.KCGame.handleMessage(KCGame.java:156)
at org.igoweb.igoweb.client.Client.connEventIn(Client.java:465)
at org.igoweb.igoweb.client.Conn.doEmit(Conn.java:268)
at org.igoweb.igoweb.client.Conn.lambda$emit$0(Conn.java:258)
at org.igoweb.igoweb.client.LockConnMutex.dispatcher(LockConnMutex.java:128)
at java.lang.Thread.run(Thread.java:748)
java.io.EOFException
at java.io.DataInputStream.readShort(DataInputStream.java:315)
at org.igoweb.igoweb.client.Conn.waitForMessages(Conn.java:302)
at java.lang.Thread.run(Thread.java:748)
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] action-value Q for unexpanded nodes

[Andy]

Thanks for letting us know the situation Aja. It must be hard for an
engineer to not be able to discuss the details of his work!

As for the first-play-urgency value, if we indulge in some reading between
the lines: It's possible to interpret the paper as saying
first-play-urgency is zero. After rereading it myself that's the way I read
it now. But if that is true maybe Aja would have said "guys the paper
already says it is zero." That would imply it's actually some other value.

That is probably reading far too much into Aja's reply, but it's something
to think about.


2017-12-06 4:47 GMT-06:00 Aja Huang :

>
>
> 2017-12-06 9:23 GMT+00:00 Gian-Carlo Pascutto :
>
>> On 03-12-17 17:57, Rémi Coulom wrote:
>> > They have a Q(s,a) term in their node-selection formula, but they
>> > don't tell what value they give to an action that has not yet been
>> > visited. Maybe Aja can tell us.
>>
>> FWIW I already asked Aja this exact question a bit after the paper came
>> out and he told me he cannot answer questions about unpublished details.
>>
>
> Yes, I did ask my manager if I could answer your question but he
> specifically said no. All I can say is that first-play-urgency is not a
> significant technical detail, and what's why we didn't specify it in the
> paper.
>
> Aja
>
>
>
>> This is not very promising regarding reproducibility considering the AZ
>> paper is even lighter on them.
>>
>> Another issue which is up in the air is whether the choice of the number
>> of playouts for the MCTS part represents an implicit balancing between
>> self-play and training speed. This is particularly relevant if the
>> evaluation step is removed. But it's possible even DeepMind doesn't know
>> the answer for sure. They had a setup, and they optimized it. It's not
>> clear which parts generalize.
>>
>> (Usually one wonders about such things in terms of algorithms, but here
>> one wonders about it in terms of hardware!)
>>
>> --
>> GCP
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] action-value Q for unexpanded nodes

[Andy]

I made a pull request to Leela, and put some data in there. It shows the
details of how Q is initialized are actually important:
https://github.com/gcp/leela-zero/pull/238


2017-12-03 19:56 GMT-06:00 Álvaro Begué :

> You are asking about the selection of the move that goes to a leaf. When
> the node before the move was expanded (in a previous playout), the value of
> Q(s,a) for that move was initialized to 0.
>
> The UCB-style formula they use in the tree part of the playout is such
> that the first few visits will follow the probability distribution from the
> policy output of the network, and over time it converges to using primarily
> the moves that have best results. So the details of how Q is initialized
> are not very relevant.
>
>
> On Sun, Dec 3, 2017 at 5:11 PM, Andy  wrote:
>
>> Álvaro, you are quoting from "Expand and evaluate (Figure 2b)". But my
>> question is about the section before that "Select (Figure 2a)". So the node
>> has not been expanded+initialized.
>>
>> As Brian Lee mentioned, his MuGo uses the parent's value, which assumes
>> without further information the value should be close to the same as before.
>>
>> LeelaZ uses 1.1 for a "first play urgency", which assumes you should
>> prioritize getting at least one evaluation from the NN for each node.
>> https://github.com/gcp/leela-zero/blob/master/src/UCTNode.cpp#L323
>>
>> Finally using a value of 0 would seem to place extra confidence in the
>> policy net values.
>>
>> I feel like MuGo's implementation makes sense, but I'm trying to get some
>> experimental evidence showing the impact before suggesting it to Leela's
>> author. So far my self-play tests with different settings do not show a big
>> impact, but I am changing other variables at the same time.
>>
>> - Andy
>>
>>
>>
>> 2017-12-03 14:30 GMT-06:00 Álvaro Begué :
>>
>>> The text in the appendix has the answer, in a paragraph titled "Expand
>>> and evaluate (Fig. 2b)":
>>>   "[...] The leaf node is expanded and and each edge (s_t, a) is
>>> initialized to {N(s_t, a) = 0, W(s_t, a) = 0, Q(s_t, a) = 0, P(s_t, a) =
>>> p_a}; [...]"
>>>
>>>
>>>
>>> On Sun, Dec 3, 2017 at 11:27 AM, Andy  wrote:
>>>
>>>> Figure 2a shows two bolded Q+U max values. The second one is going to a
>>>> leaf that doesn't exist yet, i.e. not expanded yet. Where do they get that
>>>> Q value from?
>>>>
>>>> The associated text doesn't clarify the situation: "Figure 2:
>>>> Monte-Carlo tree search in AlphaGo Zero. a Each simulation traverses the
>>>> tree by selecting the edge with maximum action-value Q, plus an upper
>>>> confidence bound U that depends on a stored prior probability P and visit
>>>> count N for that edge (which is incremented once traversed). b The leaf
>>>> node is expanded..."
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> 2017-12-03 9:44 GMT-06:00 Álvaro Begué :
>>>>
>>>>> I am not sure where in the paper you think they use Q(s,a) for a node
>>>>> s that hasn't been expanded yet. Q(s,a) is a property of an edge of the
>>>>> graph. At a leaf they only use the `value' output of the neural network.
>>>>>
>>>>> If this doesn't match your understanding of the paper, please point to
>>>>> the specific paragraph that you are having trouble with.
>>>>>
>>>>> Álvaro.
>>>>>
>>>>>
>>>>>
>>>>> On Sun, Dec 3, 2017 at 9:53 AM, Andy  wrote:
>>>>>
>>>>>> I don't see the AGZ paper explain what the mean action-value Q(s,a)
>>>>>> should be for a node that hasn't been expanded yet. The equation for 
>>>>>> Q(s,a)
>>>>>> has the term 1/N(s,a) in it because it's supposed to average over N(s,a)
>>>>>> visits. But in this case N(s,a)=0 so that won't work.
>>>>>>
>>>>>> Does anyone know how this is supposed to work? Or is it another
>>>>>> detail AGZ didn't spell out?
>>>>>>
>>>>>>
>>>>>>
>>>>>> ___
>>>>>> Computer-go mailing list
>>>>>> Computer-go@computer-go.org
>>>>>> http://computer-go.org/mailman/listinfo/computer-go
>>>>>>
>>>>>
>>>>>
>>>>> ___
>>>>> Computer-go mailing list
>>>>> Computer-go@computer-go.org
>>>>> http://computer-go.org/mailman/listinfo/computer-go
>>>>>
>>>>
>>>>
>>>> ___
>>>> Computer-go mailing list
>>>> Computer-go@computer-go.org
>>>> http://computer-go.org/mailman/listinfo/computer-go
>>>>
>>>
>>>
>>> ___
>>> Computer-go mailing list
>>> Computer-go@computer-go.org
>>> http://computer-go.org/mailman/listinfo/computer-go
>>>
>>
>>
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] action-value Q for unexpanded nodes

[Andy]

Álvaro, you are quoting from "Expand and evaluate (Figure 2b)". But my
question is about the section before that "Select (Figure 2a)". So the node
has not been expanded+initialized.

As Brian Lee mentioned, his MuGo uses the parent's value, which assumes
without further information the value should be close to the same as before.

LeelaZ uses 1.1 for a "first play urgency", which assumes you should
prioritize getting at least one evaluation from the NN for each node.
https://github.com/gcp/leela-zero/blob/master/src/UCTNode.cpp#L323

Finally using a value of 0 would seem to place extra confidence in the
policy net values.

I feel like MuGo's implementation makes sense, but I'm trying to get some
experimental evidence showing the impact before suggesting it to Leela's
author. So far my self-play tests with different settings do not show a big
impact, but I am changing other variables at the same time.

- Andy



2017-12-03 14:30 GMT-06:00 Álvaro Begué :

> The text in the appendix has the answer, in a paragraph titled "Expand and
> evaluate (Fig. 2b)":
>   "[...] The leaf node is expanded and and each edge (s_t, a) is
> initialized to {N(s_t, a) = 0, W(s_t, a) = 0, Q(s_t, a) = 0, P(s_t, a) =
> p_a}; [...]"
>
>
>
> On Sun, Dec 3, 2017 at 11:27 AM, Andy  wrote:
>
>> Figure 2a shows two bolded Q+U max values. The second one is going to a
>> leaf that doesn't exist yet, i.e. not expanded yet. Where do they get that
>> Q value from?
>>
>> The associated text doesn't clarify the situation: "Figure 2: Monte-Carlo
>> tree search in AlphaGo Zero. a Each simulation traverses the tree by
>> selecting the edge with maximum action-value Q, plus an upper confidence
>> bound U that depends on a stored prior probability P and visit count N for
>> that edge (which is incremented once traversed). b The leaf node is
>> expanded..."
>>
>>
>>
>>
>>
>>
>> 2017-12-03 9:44 GMT-06:00 Álvaro Begué :
>>
>>> I am not sure where in the paper you think they use Q(s,a) for a node s
>>> that hasn't been expanded yet. Q(s,a) is a property of an edge of the
>>> graph. At a leaf they only use the `value' output of the neural network.
>>>
>>> If this doesn't match your understanding of the paper, please point to
>>> the specific paragraph that you are having trouble with.
>>>
>>> Álvaro.
>>>
>>>
>>>
>>> On Sun, Dec 3, 2017 at 9:53 AM, Andy  wrote:
>>>
>>>> I don't see the AGZ paper explain what the mean action-value Q(s,a)
>>>> should be for a node that hasn't been expanded yet. The equation for Q(s,a)
>>>> has the term 1/N(s,a) in it because it's supposed to average over N(s,a)
>>>> visits. But in this case N(s,a)=0 so that won't work.
>>>>
>>>> Does anyone know how this is supposed to work? Or is it another detail
>>>> AGZ didn't spell out?
>>>>
>>>>
>>>>
>>>> ___
>>>> Computer-go mailing list
>>>> Computer-go@computer-go.org
>>>> http://computer-go.org/mailman/listinfo/computer-go
>>>>
>>>
>>>
>>> ___
>>> Computer-go mailing list
>>> Computer-go@computer-go.org
>>> http://computer-go.org/mailman/listinfo/computer-go
>>>
>>
>>
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] action-value Q for unexpanded nodes

[Andy]

Figure 2a shows two bolded Q+U max values. The second one is going to a
leaf that doesn't exist yet, i.e. not expanded yet. Where do they get that
Q value from?

The associated text doesn't clarify the situation: "Figure 2: Monte-Carlo
tree search in AlphaGo Zero. a Each simulation traverses the tree by
selecting the edge with maximum action-value Q, plus an upper confidence
bound U that depends on a stored prior probability P and visit count N for
that edge (which is incremented once traversed). b The leaf node is
expanded..."






2017-12-03 9:44 GMT-06:00 Álvaro Begué :

> I am not sure where in the paper you think they use Q(s,a) for a node s
> that hasn't been expanded yet. Q(s,a) is a property of an edge of the
> graph. At a leaf they only use the `value' output of the neural network.
>
> If this doesn't match your understanding of the paper, please point to the
> specific paragraph that you are having trouble with.
>
> Álvaro.
>
>
>
> On Sun, Dec 3, 2017 at 9:53 AM, Andy  wrote:
>
>> I don't see the AGZ paper explain what the mean action-value Q(s,a)
>> should be for a node that hasn't been expanded yet. The equation for Q(s,a)
>> has the term 1/N(s,a) in it because it's supposed to average over N(s,a)
>> visits. But in this case N(s,a)=0 so that won't work.
>>
>> Does anyone know how this is supposed to work? Or is it another detail
>> AGZ didn't spell out?
>>
>>
>>
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] action-value Q for unexpanded nodes

[Andy]

I don't see the AGZ paper explain what the mean action-value Q(s,a) should
be for a node that hasn't been expanded yet. The equation for Q(s,a) has
the term 1/N(s,a) in it because it's supposed to average over N(s,a)
visits. But in this case N(s,a)=0 so that won't work.

Does anyone know how this is supposed to work? Or is it another detail AGZ
didn't spell out?
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Significance of resignation in AGZ

[Andy]

Brian, do you have any experiments showing what kind of impact it has? It
sounds like you have tried both with and without your ad hoc first pass
approach?




2017-12-01 15:29 GMT-06:00 Brian Sheppard via Computer-go <
computer-go@computer-go.org>:

> I have concluded that AGZ's policy of resigning "lost" games early is
> somewhat significant. Not as significant as using residual networks, for
> sure, but you wouldn't want to go without these advantages.
>
> The benefit cited in the paper is speed. Certainly a factor. I see two
> other advantages.
>
> First is that training does not include the "fill in" portion of the game,
> where every move is low value. I see a specific effect on the move ordering
> system, since it is based on frequency. By eliminating training on
> fill-ins, the prioritization function will not be biased toward moves that
> are not relevant to strong play. (That is, there are a lot of fill-in
> moves, which are usually not best in the interesting portion of the game,
> but occur a lot if the game is played out to the end, and therefore the
> move prioritization system would predict them more often.) My ad hoc
> alternative is to not train on positions after the first pass in a game.
> (Note that this does not qualify as "zero knowledge", but that is OK with
> me since I am not trying to reproduce AGZ.)
>
> Second is the positional evaluation is not training on situations where
> everything is decided, so less of the NN capacity is devoted to situations
> in which nothing can be gained.
>
> As always, YMMV.
>
> Best,
> Brian
>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Is MCTS needed?

[Andy]

Matthew Lai's paper "Giraffe: Using Deep Reinforcement Learning to Play
Chess" has a chapter called "Probabilistic Search". That should be directly
applicable?

https://arxiv.org/abs/1509.01549

- Andy


2017-11-16 9:43 GMT-06:00 Petr Baudis :

>   Hi,
>
>   when explaining AlphaGo Zero to a machine learning audience yesterday
>
> (https://docs.google.com/presentation/d/
> 1VIueYgFciGr9pxiGmoQyUQ088Ca4ouvEFDPoWpRO4oQ/view)
>
> it occurred to me that using MCTS in this setup is actually such
> a kludge!
>
>   Originally, we used MCTS because with the repeated simulations,
> we would be improving the accuracy of the arm reward estimates.  MCTS
> policies assume stationary distributions, which is violated every time
> we expand the tree, but it's an okay tradeoff if all you feed into the
> tree are rewards in the form of just Bernoulli trials.  Moreover, you
> could argue evaluations are somewhat monotonic with increasing node
> depths as you are basically just fixing a growing prefix of the MC
> simulation.
>
>   But now, we expand the nodes literally all the time, breaking the
> stationarity possibly in drastic ways.  There are no reevaluations that
> would improve your estimate.  The input isn't binary but an estimate in
> a continuous space.  Suddenly the Multi-armed Bandit analogy loses a lot
> of ground.
>
>   Therefore, can't we take the next step, and do away with MCTS?  Is
> there a theoretical viewpoint from which it still makes sense as the best
> policy improvement operator?
>
>   What would you say is the current state-of-art game tree search for
> chess?  That's a very unfamiliar world for me, to be honest all I really
> know is MCTS...
>
> --
> Petr Baudis, Rossum
> Run before you walk! Fly before you crawl! Keep moving forward!
> If we fail, I'd rather fail really hugely.  -- Moist von Lipwig
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] AlphaGo Zero

[Andy]

I agree with your main point that the first batch of games will be totally
random moves. I just wanted to make a small point that even for totally
random play, the network should be able to learn something about mid-game
positions as well. At move 100, a position with 50 white stones and 40
black stones is likely to be a win for white, even with completely random
play from there, since white has captured 10 black stones.


2017-10-26 8:17 GMT-05:00 Gian-Carlo Pascutto :

> On 25-10-17 16:00, Petr Baudis wrote:
> > That makes sense.  I still hope that with a much more aggressive
> > training schedule we could train a reasonable Go player, perhaps at
> > the expense of worse scaling at very high elos...  (At least I feel
> > optimistic after discovering a stupid bug in my code.)
>
> By the way, a trivial observation: the initial network is random, so
> there's no point in using it for playing the first batch of games. It
> won't do anything useful until it has run a learning pass on a bunch of
> "win/loss" scored games and it can at least tell who is the likely
> winner in the final position (even if it mostly won't be able to make
> territory at first).
>
> This suggests that bootstrapping probably wants 500k starting games with
> just random moves.
>
> FWIW, it does not seem easy to get the value part of the network to
> converge in the dual-res architecture, even when taking the appropriate
> steps (1% weighting on error, strong regularizer).
>
> --
> GCP
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Source code (Was: Reducing network size? (Was: AlphaGo Zero))

[Andy]

Gian-Carlo, I didn't realize at first that you were planning to create a
crowd-sourced project. I hope this project can get off the ground and
running!

I'll look into installing this but I always find it hard to get all the
tool chain stuff going.



2017-10-24 15:02 GMT-05:00 Gian-Carlo Pascutto :

> On 23-10-17 10:39, Darren Cook wrote:
> >> The source of AlphaGo Zero is really of zero interest (pun intended).
> >
> > The source code is the first-hand account of how it works, whereas an
> > academic paper is a second-hand account. So, definitely not zero use.
>
> This should be fairly accurate:
>
> https://github.com/gcp/leela-zero
>
> --
> GCP
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] it's alphago

[Andy]

What is Ray? Strongest open source bot? Anyone have a link to it?

On Fri, Jan 6, 2017 at 3:39 AM, Hiroshi Yamashita  wrote:

> If value net is the most important part for over pro level, the problem is
> making strong selfplay games.
>
> 1. make 30 million selfplay games.
> 2. make value net.
> 3. use this value net for selfplay program.
> 4. go to (1)
>
> I don't know when the progress will stop by this loop.
> But if once strong enough selfplay games are published, everyone can make
> pro level program.
> 30 million is big number. It needs many computers.
> Computer Go community may be able to share this work.
> I can offer Aya, it is not open-source though. Maybe Ray(strongest open
> source so far)  is better choice.
>
> Thanks,
> Hiroshi Yamashita
>
> - Original Message - From: 
> To: 
> Sent: Friday, January 06, 2017 4:50 PM
> Subject: Re: [Computer-go] it's alphago
>
>
> Competitive with Alpha-go, one developer, not possible. I do think it is
> possible to make a pro level program with one person or a small team. Look
> at Deep Zen and Aya for example. I expect I’ll get there (pro level) with
> Many Faces as well.
>
> David
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] DCNN can solve semeai?

[Andy]

The Deepmind paper has a short section on the Rollout Policy they use, it
looks like they made some improvements on for their rollouts, maybe they
are better at handling semeai than previous methods. The response and
non-response patterns sound similar, but they also include liberty counts.
I don't remember that being included in classic Mogo 3x3 patterns.

They also talk about caching moves from the search tree. So if the tree has
already found the correct answers to sente moves in a semeai, they could be
applied in the rollout as well.

Has anyone tried techniques similar to these?





On Tue, Feb 2, 2016 at 2:02 PM, Hiroshi Yamashita  wrote:

> Hi Michael,
>
> It's an intersting idea.
> Maybe I could collect many LD positions from pro games.
> So LD and semeai solver will be needed. Or just use LD
> problems with sequence.
> Without difficult feature, DCNN may find answer.
>
> Regards,
> Hiroshi Yamashita
>
> - Original Message - From: "Michael Sué" 
> To: 
> Sent: Wednesday, February 03, 2016 4:39 AM
> Subject: Re: [Computer-go] DCNN can solve semeai?
>
>
> Hi,
>>
>> I would expect this to happen if the system is trained by normal games,
>> only. But I think the system should see actual live-and-death (LD)
>> sequences (from some collection) to be able to learn about them and not
>> soak this knowledge up from a whole game where most of the moves are
>> "noise" compared to what you ask it to do later.
>> So the training data could be half and half normal games and LD sequences.
>>
>> - Michael.
>>
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Facebook Go AI

[Andy]

As of about an hour ago darkforest and darkfores1 have started playing
rated games on KGS!


2015-11-23 11:28 GMT-06:00 Andy :

> So the KGS bots darkforest and darkfores1 play with only DCNN, no MCTS
> search added? I wish they would put darkfores2 with MCTS on KGS, why not
> put your strongest bot out there?
>
>
>
>
> 2015-11-23 10:38 GMT-06:00 Petr Baudis :
>
>> The numbers look pretty impressive! So this DNN is as strong as
>> a full-fledged MCTS engine with non-trivial thinking time. The increased
>> supervision is a nice idea, but even barring that this seems like quite
>> a boost to the previously published results?  Surprising that this is
>> just thanks to relatively simple tweaks to representations and removing
>> features... (Or is there anything important I missed?)
>>
>> I'm not sure what's the implementation difference between darkfores1 and
>> darkfores2, it's a bit light on detail given how huge the winrate delta
>> is, isn't it? ("we fine-tuned the learning rate")  Hopefully peer review
>> will help.
>>
>> Do I understand it right that in the tree, they sort moves by their
>> probability estimate, keep only moves whose probability sum up to
>> 0.8, prune the rest and use just plain UCT with no priors afterwards?
>> The result with +MCTS isn't at all convincing - it just shows that
>> MCTS helps strength, which isn't so surprising, but the extra thinking
>> time spent corresponds to about 10k->150k playouts increase in Pachi,
>> which may not be a good trade for +27/4.5/1.2% winrate increase.
>>
>> On Mon, Nov 23, 2015 at 09:54:37AM +0100, Rémi Coulom wrote:
>> > It is darkforest, indeed:
>> >
>> > Title: Better Computer Go Player with Neural Network and Long-term
>> > Prediction
>> >
>> > Authors: Yuandong Tian, Yan Zhu
>> >
>> > Abstract:
>> > Competing with top human players in the ancient game of Go has been a
>> > long-term goal of artificial intelligence. Go's high branching factor
>> makes
>> > traditional search techniques ineffective, even on leading-edge
>> hardware,
>> > and Go's evaluation function could change drastically with one stone
>> change.
>> > Recent works [Maddison et al. (2015); Clark & Storkey (2015)] show that
>> > search is not strictly necessary for machine Go players. A pure
>> > pattern-matching approach, based on a Deep Convolutional Neural Network
>> > (DCNN) that predicts the next move, can perform as well as Monte Carlo
>> Tree
>> > Search (MCTS)-based open source Go engines such as Pachi [Baudis &
>> Gailly
>> > (2012)] if its search budget is limited. We extend this idea in our bot
>> > named darkforest, which relies on a DCNN designed for long-term
>> predictions.
>> > Darkforest substantially improves the win rate for pattern-matching
>> > approaches against MCTS-based approaches, even with looser search
>> budgets.
>> > Against human players, darkforest achieves a stable 1d-2d level on KGS
>> Go
>> > Server, estimated from free games against human players. This
>> substantially
>> > improves the estimated rankings reported in Clark & Storkey (2015),
>> where
>> > DCNN-based bots are estimated at 4k-5k level based on performance
>> against
>> > other machine players. Adding MCTS to darkforest creates a much stronger
>> > player: with only 1000 rollouts, darkforest+MCTS beats pure darkforest
>> 90%
>> > of the time; with 5000 rollouts, our best model plus MCTS beats Pachi
>> with
>> > 10,000 rollouts 95.5% of the time.
>> >
>> > http://arxiv.org/abs/1511.06410
>>
>> --
>> Petr Baudis
>> If you have good ideas, good data and fast computers,
>> you can do almost anything. -- Geoffrey Hinton
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>
>
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Facebook Go AI

[Andy]

So the KGS bots darkforest and darkfores1 play with only DCNN, no MCTS
search added? I wish they would put darkfores2 with MCTS on KGS, why not
put your strongest bot out there?




2015-11-23 10:38 GMT-06:00 Petr Baudis :

> The numbers look pretty impressive! So this DNN is as strong as
> a full-fledged MCTS engine with non-trivial thinking time. The increased
> supervision is a nice idea, but even barring that this seems like quite
> a boost to the previously published results?  Surprising that this is
> just thanks to relatively simple tweaks to representations and removing
> features... (Or is there anything important I missed?)
>
> I'm not sure what's the implementation difference between darkfores1 and
> darkfores2, it's a bit light on detail given how huge the winrate delta
> is, isn't it? ("we fine-tuned the learning rate")  Hopefully peer review
> will help.
>
> Do I understand it right that in the tree, they sort moves by their
> probability estimate, keep only moves whose probability sum up to
> 0.8, prune the rest and use just plain UCT with no priors afterwards?
> The result with +MCTS isn't at all convincing - it just shows that
> MCTS helps strength, which isn't so surprising, but the extra thinking
> time spent corresponds to about 10k->150k playouts increase in Pachi,
> which may not be a good trade for +27/4.5/1.2% winrate increase.
>
> On Mon, Nov 23, 2015 at 09:54:37AM +0100, Rémi Coulom wrote:
> > It is darkforest, indeed:
> >
> > Title: Better Computer Go Player with Neural Network and Long-term
> > Prediction
> >
> > Authors: Yuandong Tian, Yan Zhu
> >
> > Abstract:
> > Competing with top human players in the ancient game of Go has been a
> > long-term goal of artificial intelligence. Go's high branching factor
> makes
> > traditional search techniques ineffective, even on leading-edge hardware,
> > and Go's evaluation function could change drastically with one stone
> change.
> > Recent works [Maddison et al. (2015); Clark & Storkey (2015)] show that
> > search is not strictly necessary for machine Go players. A pure
> > pattern-matching approach, based on a Deep Convolutional Neural Network
> > (DCNN) that predicts the next move, can perform as well as Monte Carlo
> Tree
> > Search (MCTS)-based open source Go engines such as Pachi [Baudis & Gailly
> > (2012)] if its search budget is limited. We extend this idea in our bot
> > named darkforest, which relies on a DCNN designed for long-term
> predictions.
> > Darkforest substantially improves the win rate for pattern-matching
> > approaches against MCTS-based approaches, even with looser search
> budgets.
> > Against human players, darkforest achieves a stable 1d-2d level on KGS Go
> > Server, estimated from free games against human players. This
> substantially
> > improves the estimated rankings reported in Clark & Storkey (2015), where
> > DCNN-based bots are estimated at 4k-5k level based on performance against
> > other machine players. Adding MCTS to darkforest creates a much stronger
> > player: with only 1000 rollouts, darkforest+MCTS beats pure darkforest
> 90%
> > of the time; with 5000 rollouts, our best model plus MCTS beats Pachi
> with
> > 10,000 rollouts 95.5% of the time.
> >
> > http://arxiv.org/abs/1511.06410
>
> --
> Petr Baudis
> If you have good ideas, good data and fast computers,
> you can do almost anything. -- Geoffrey Hinton
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Building A Computer Go AI

[Andy]

Here is a simple working implementation.
https://github.com/pasky/michi

>From the beginning of the readme:

Michi --- Minimalistic Go MCTS Engine

Michi aims to be a minimalistic but full-fledged Computer Go program based
on state-of-art methods (Monte Carlo Tree Search) and written in Python.
Our goal is to make it easier for new people to enter the domain of
Computer Go, peek under the hood of a "real" playing engine and be able to
learn by hassle-free experiments - with the algorithms, add heuristics, etc.

The algorithm code size is 540 lines of code (without user interface,
tables and empty lines / comments). Currently, it can often win against
GNUGo on 9×9 on an old i3 notebook, be about even with GNUGo on 15×15 on a
modern higher end computer and about two stones weaker on 19×19 (spending
no more than 30s per move).

This is not meant to be a competitive engine; simplicity and clear code is
preferred over optimization (after all, it's in Python!). But compared to
other minimalistic engines, this one should be able to beat beginner
intermediate human players, and I believe that a *fast* implementation of
exactly the same heuristics would be around 4k KGS or even better.

Michi is distributed under the MIT licence. Now go forth, hack and peruse!





On Fri, Aug 21, 2015 at 7:06 AM, robertfinkng...@o2.co.uk <
robertfinkng...@o2.co.uk> wrote:

> Hi,
>
> Good news. There are a variety of open source projects out there,
> including both complete programs (Fuego, Pachi) as well as libraries to
> build your own Go engine (libEgo). There are also a wealth of papers
> explaining the theory behind the top algorithms. Try googling "AMAF
> algorithm" or "RAVE algorithm" or "MCTS algorithm" or "TD Search algorithm"
> as a starting point. There is a nice Thesis on Pachi too google "Pachi
> Thesis".
>
> I hope this helps :-)
>
> Regards
>
> Raffles
>
>
> On 21-Aug-15 8:48, CaiGengYang wrote:
>
>> Hello …
>>
>>
>> I am a 3d~~5d go player from Singapore.
>>
>> Keen to learn how to build a powerful Computer Go AI to compete in the
>> Computer Go Tournament and also for admissions to a Computer Science
>> college program.
>>
>> Have very little programming experience except following some code
>> examples on CodeAcademy … how do I start building a Computer Go AI ?
>>
>>
>> Gengyang
>> ___
>> Computer-go mailing list
>> Computer-go@computer-go.org
>> http://computer-go.org/mailman/listinfo/computer-go
>>
>> -
>> No virus found in this message.
>> Checked by AVG - www.avg.com
>> Version: 2015.0.6125 / Virus Database: 4392/10476 - Release Date: 08/21/15
>>
>
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [computer-go] Big trees

[Andy]

Put it on cgos and see how good it is!

On Fri, Jul 10, 2009 at 9:10 PM, Michael Williams <
michaelwilliam...@gmail.com> wrote:

> Now that I have this system of generating really big game trees, what sort
> of interesting things could I do with it?  The exact number of nodes I can
> store is not exact because I'm doing various things to reduce each node's
> footprint when it goes to disk.  I'm currently building a tree that is bushy
> at the root (heavy exploration term) and normal UCT beneath that.  It is at
> 28 billion nodes now and projecting a capacity of 122 billion.  The current
> node rate is about 130k per second (on 1 Core2 core).  This is on a 9x9
> board.  I'm still using Libego for playouts.  And I'm deleting
> symmetrically-equivalent moves from the tree.  That is all that gets pruned.
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] Extra komi for Handicap stone compensation on KGS

[Andy]

FYI, here is a link and the relevant quotes for the way KGS gives
white extra komi based on handicap stones in some rules.

http://www.gokgs.com/help/rulesets.html

Chinese - "The white player is given one point extra komi for every
handicap stone that Black gets at the start of the game."

AGA - "...the white player is given one point of extra komi for all
but the first handicap stone that Black gets at the start of the game"

New Zealand - "New Zealand rules are like Chinese rules except that
playing a multi-stone suicide move is allowed."  (I assume this means
New Zealand also includes extra komi for handicap stones).
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] The Zen program

[Andy]

Yamato,

It looks like Zen19 doesn't implement handicap stone komi compensation
the same way kgs does for Chinese rules.  It's the only reason I can
think that it lost this game:

http://files.gokgs.com/games/2009/3/27/TaPaHka-Zen19.sgf

- Andy, aka yoyoma


On Thu, Mar 26, 2009 at 5:54 PM, Yamato  wrote:
> Rémi Coulom wrote:
>>Martin Mueller wrote:
>>> Zen has been getting very impressive results on CGOS. Yamato-san,
>>> could you tell us a little bit about yourself and your program?
>>>
>>> Thanks
>>>
>>>     Martin
>>
>>Yes, very strong results. Congratulations to Yamato-san for beating
>>Crazy Stone.
>>
>>I am curious about hardware. On CGOS, Crazy Stone is running on core2
>>duo (2 threads).
>>
>>Rémi
>
> Hi Martin & Rémi, thanks for having interest in my program.
> Zen runs on Quad-Core AMD Opteron. I suppose it has nearly the same
> strength as Crazy Stone. It is running on KGS now and you can see its
> games from here.
> http://www.gokgs.com/gameArchives.jsp?user=zen19
>
> I am an individual programmer in Japan. Any questions are welcome.
>
> --
> Yamato
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Congratulations to David Fotland and Many Faces

[Andy]

On Mon, Feb 16, 2009 at 7:45 PM, Andy  wrote:
> See attached a copy of the .sgf.  It was played private on KGS so you
> can't get it there directly.  One of the admins cloned it and I saved
> it off locally.
>
> I changed the result to be B+4.5 instead of W+2.5.

I forgot to make a disclaimer:  I am not an organizer of the event or
a member of the CrazyStone team.  I just happened to see a copy of the
game, and I changed the result based on some reports I saw.  I've
received a few questions about this but you should really direct any
questions to an authority.

- Andy
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Congratulations to David Fotland and Many Faces

[Andy]

See attached a copy of the .sgf.  It was played private on KGS so you
can't get it there directly.  One of the admins cloned it and I saved
it off locally.

I changed the result to be B+4.5 instead of W+2.5.


On Sat, Feb 14, 2009 at 2:18 PM, Bob Hearn  wrote:
> Many Faces won its match today against James Kerwin 1p, played at 7 stones,
> by 4.5 points. The program was running on a 32-core cluster supplied by
> Microsoft. The match was played live in front of a press briefing at the
> 2009 AAAS general meeting.
>
> The game record will be available on KGS shortly, if it is not already. Note
> that KGS scored the game incorrectly, so the result in the game record is
> wrong.
>
> As David says:
>
>> This version of Many Faces plays pure Chinese rules, so it will place
>> handicaps in nonstandard places.  It counts all stones on the board at the
>> end of the games as points, including the original handicap stones.
>>
>> KGS scores differently.  In a Chinese handicap game it gives one point
>> compensation to white for each handicap stone.
>>
>> We need to play with pure Chinese rules.  Keep this in mind when setting
>> the
>> handicap.  Each handicap stone is worth one point more than a Japanese
>> handicap stone.
>>
>> At the end of the game, the KGS score will be incorrect.  It will show
>> white
>> with one extra point for each handicap stone.
>
>
> Thanks to both David and Jim for participating. I got a lot of good
> questions at the AAAS press conference; hopefully this will add some popular
> interest.
>
> Bob Hearn
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>


JKerwin-ManyFaces1.sgf
Description: application/go-sgf
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: Opportunity to promote ...

[Andy]

On Wed, Nov 19, 2008 at 12:24 PM, Christoph Birk <[EMAIL PROTECTED]> wrote:

> On Wed, 19 Nov 2008, [EMAIL PROTECTED] wrote:
>
>> I think that would not be enough, because that would only fix one point.
>>
>
> You can use the width too. That should give a pretty good comparision
> for moderatly strong/weak players (see below).
>
>  EGF ratings are not pure Elo ratings. EGF ratings are weighted to fit 100
>> points for one handicap stone, which happens to match about 65% winning
>> percentage in even games for medium level players (around 3k).
>>
>
> That should not matter much. The typical chess player should be
> "as strong" as the typical Go player and I also expect the strength
> distribution to follow similar lines.
>
>  Also, I am not aware that there exists a histogram of the worldwide go
>> population.
>>
>
> Why would you need world-wide data? Use US-Go/Chess or European data.
> The lack of pros in this distributions should not matter much, as these
> are very few at the top end of the distribution.
>
>
> Christoph
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>


Here you go!

http://senseis.xmp.net/?RatingHistogramComparisons

  %   AGA  KGS  USCFEGF
 ---   | -- | -- | - | - |
  1%   | -34.61 | -24.26 |   444 |   100 |
  2%   | -32.58 | -22.30 |   531 |   100 |
  5%   | -27.69 | -19.20 |   663 |   153 |
 10%   | -23.47 | -15.36 |   793 |   456 |
 20%   | -18.54 | -11.26 |   964 |   953 |
 30%   | -13.91 |  -8.94 |  1122 |  1200 |
 40%   |  -9.90 |  -7.18 |  1269 |  1387 |
 50%   |  -7.10 |  -5.65 |  1411 |  1557 |
 60%   |  -4.59 |  -4.19 |  1538 |  1709 |
 70%   |  -1.85 |  -2.73 |  1667 |  1884 |
 80%   |   2.10 |  -1.28 |  1807 |  2039 |
 90%   |   4.71 |   2.52 |  1990 |  2217 |
 95%   |   6.12 |   3.88 |  2124 |  2339 |
 98%   |   7.41 |   5.29 |  2265 |  2460 |
 99%   |   8.15 |   6.09 |  2357 |  2536 |
 99.5% |   8.70 |   7.20 |  2470 |  2604 |
 99.9% |   9.64 |   pro  |  2643 |  2747 |
 top   |  10.12 |9p  |  2789 |  2809 |
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] Time controls in bots vs human matches

[Andy]

I think for bot vs human, the time control should include byoyomi/overtime
of some kind instead of sudden death.  I'm afraid in one of these exhibition
matches the human will be winning but lose on time.  It would be especially
bad if the bot was playing meaningless invasions or territory filling moves
when the human loses.

Is there a big reason not to use some overtime?  It could be relatively
quick, say 10s per move.  Just to prevent losing a won game on time.  Some
people would argue the human should manage his time better, but people are
much more used to playing with overtime, and accommodating this doesn't seem
like a big deal.  Even on CGOS there is a 1s Bronstein delay to prevent
silly time loses due to lag.  We need the same thing for humans except that
for humans it needs to be a bit more than 1s.

Same thing for bots on KGS.

- Andy
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Kaori-Crazystone

[Andy]

On Thu, Sep 4, 2008 at 11:09 AM, Rémi Coulom <[EMAIL PROTECTED]>wrote:

> Andy wrote:
>
>> I'm excited to see a computer reach 1d as well.  For me I'm waiting to see
>> a bot hold a 1d rating consistently on kgs.  Right now CrazyStone has been
>> rated 1d briefly, but hasn't been able to maintain it.  It's currently 1k.
>>
>> I put a small table of the progress of a few bot's ratings on kgs at
>> http://senseis.xmp.net/?KGSBotRatings
>>
>> I would like to see MogoTiTan play many rated games on KGS and see how it
>> does there.  Anyone have a few million dollars lying around to sponsor this?
>>  :)
>>
>
> Leela is becoming strong. It has reached 1k now.
>
> Rémi
>

Thanks, I have updated http://senseis.xmp.net/?KGSBotRatings

- Andy
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Kaori-Crazystone

[Andy]

I'm excited to see a computer reach 1d as well.  For me I'm waiting to see a
bot hold a 1d rating consistently on kgs.  Right now CrazyStone has been
rated 1d briefly, but hasn't been able to maintain it.  It's currently 1k.

I put a small table of the progress of a few bot's ratings on kgs at
http://senseis.xmp.net/?KGSBotRatings

I would like to see MogoTiTan play many rated games on KGS and see how it
does there.  Anyone have a few million dollars lying around to sponsor
this?  :)


On Thu, Sep 4, 2008 at 10:48 AM, Don Dailey <[EMAIL PROTECTED]> wrote:

> I meant to add that we cannot calculate an upper bound on it's strength
> since there was only 1 game and it was a win.
>
> What I'm trying to determine is if we can say with a high degree of
> confidence yet that computers have achieved the 1 dan level?   This has
> been kind of a holy grail of computer go in my opinion - even if it
> wasn't directly articulated (or perhaps it was?)
>
> - Don
>
>
> On Thu, 2008-09-04 at 11:38 -0400, Don Dailey wrote:
> > It's difficult for me to understand this due to different ranking
> > systems and pro ratings vs amateur ratings.   I see here listed as a 4
> > dan player on this page:
> >
> > http://www.nihonkiin.or.jp/player/htm/ki000343.htm
> >
> >
> > Is that 4 dan pro?  My understanding is something like this:
> >
> > kyu player are casual players (or weak tournament players)
> >
> > low dan players are something like advanced amateurs or experts and weak
> > masters in chess.
> >
> > Pro's are like super high dan players and there is not very much
> > difference between ranks compared to regular dan players.  I have heard
> > that a 1d professional will beat a 9d professional with 3 or 4 stones.
> >
> > So a 1d pro is something like a 7 or 8d+ amateur?
> >
> > Is this all "roughly" correct?
> >
> > So I assume that Aoba Kaori is a 4d professional?  That would relate to
> > something in the ballpark of 9 or 10d amateur if there were such a
> > thing.   And with 8 stones handicap, this implies that CrazyStone did
> > what a 2d+ would have done,  or it is weaker than 2d but got lucky.  So
> > it's "performance rating" for that one game is lower bounded at around 1
> > or 2 dan.   Since it won the game we could pick 2 dan as a better lower
> > bound guess although since it won we do not have a reasonable upper
> > bound guess on it's performance except our own credulity.
> >
> > Does what I said make any sense?  I am not a go player and I'm not very
> > comfortable with this guesswork.   In chess, if you beat a player I am
> > used to thinking in terms of setting a performance rating of around 400
> > ELO higher for that one game.   I know this is not precise, but I also
> > think of 400 ELO subtracted from the player you beat as a kind of
> > "estimated" lower bound on your strength.  If you beat a 2500 ELO chess
> > player, it's a relatively safe bet that you are at least 2100 ELO in
> > strength although technically there is a chance you could lose to
> > anybody, even a random move generator.
> >
> > I know this isn't precise language, but how many ranks would give us
> > around 90 - 95% confidence of superiority?If I beat a 5 dan player,
> > could you say that it's "very likely" I am at least 3 dan in strength?
> >
> > I'm thinking that if we estimate Aoba at 10d amateur and CrazyStone wins
> > with 8 stone handicap, it is roughly equivalent to beating a 2d player
> > without handicap and that we can subtract 2 stones to say that with
> > pretty high confidence CrazyStone is playing at least 1 kyu  (but that's
> > it's much more likely Crazy Stone is stronger than this - after all it
> > performed in this one game at least as well as 2d player.)
> >
> >
> > - Don
> >
> >
> >
> >
> > On Thu, 2008-09-04 at 16:28 +0200, Rémi Coulom wrote:
> > > terry mcintyre wrote:
> > > > Congratulations!
> > > >
> > >
> > > Thanks.
> > >
> > > > I'm dying for details! What was the time limit?
> > >
> > > The organizers asked that the program should play at a constant time
> (30
> > > second) per move. The sgf file contains time stamps (you can see the
> > > time with gogui, for instance). I don't know what was her time control,
> > > but she apparently played at the same pace as the program.
> > >
> > > >  Did the game end on time or by resignation at move 179?
> > > >
> > >
> > > She resigned.
> > >
> > > > The pro was Aoba Kaori, yes?
> > > >
> > >
> > > Yes.
> > >
> > > The only other information I have about the match are these pages in
> > > Japanese:
> > >
> https://secure1.gakkai-web.net/gakkai/fit/program/html/event/event.html#6
> > > http://www.ipsj.or.jp/10jigyo/fit/fit2008/events.html#1-4-1
> > >
> > > I hope the organizers can send me some photos tomorrow. Then I will set
> > > up a web page and tell the list.
> > >
> > > Rémi
> > > ___
> > > computer-go mailing list
> > > computer-go@computer-go.org
> > > http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] yet a mogo vs human game

[Andy]

On Wed, Aug 27, 2008 at 5:45 PM, Rémi Coulom <[EMAIL PROTECTED]>wrote:

> Don Dailey wrote:
>
>>
>> Yes, I believe it does generalize on average.
>> This data matches my 13x13 study pretty closely,  about 62% give or take
>> for each doubling. That is about 90 ELO or so.   I have heard that
>> 100 ELO is 1 stone which is what I was basing this on.   But it's not
>> clear to me at all if that is true.   So I can only guess that 4x in
>> Mogo is worth something like 1 or 2 stones or something between.
>> - Don
>>
>
> According to my experience with Go data, it is not possible to give the
> value of one stone in terms of Elo ratings. For weak players, one stone is a
> lot less than 100 Elo. For stronger players, it may be more.
>
> Also, it is very important to understand that the Elo model is very wrong,
> and Elo against humans has nothing to do with Elo against computers (and
> even less with Elo against the previous version). In games against GNU Go,
> Crazy Stone improved 200-300 Elo points in one year. On KGS, this translated
> into an improvement from 2k to 1k.
>
> Rémi


I made a relevant post about this awhile back, below I pasted it.  Maybe
I'll make a senseis page for this.  KGS varies the "k" factor according to
the strength of the players to account for the differences in winning
probabilities in even games for weak players vs strong players.  The EGF
formula has a similar factor "a" that varies for the same reason.  So the
EGF Elo ratings which set 1 stone = 100 points doesn't correspond directly
to the classic Elo probabilities.

So for KGS, improving one stone can be anywhere from 139 to 226 classic Elo
points.
For CrazyStone's improvement from 2k to 1k, this represents around a 182
classic Elo points improvement.

-
See below I created a table that shows the transformation from KGS ratings
to the Elo that CGOS uses.  I set 6k=1800 because I believe that is what GNU
3.7.10 is under both systems.  Does anyone have more data points for bots
that play on both systems?

Also is there an "all times" list for 19x19?  CS-9-17-2CPU is on top at
2297, but I don't think that's the strongest CrazyStone.  I think other strong
19x19 bots are not on the current list either.

To create the table I calculated the probability for a player of rank X to
upset a player of rank X+1.  Then I converted this to the equivalent number
of Elo points for the same upset rate.  Finally I made a running total,
setting 6k to 1800.  The transform could be done using calculus I suppose
but it's been a long time since I did that!  :)


   ELO  set
rank k1 rank  6k=1800
 10k   0.80139 1,244
  9k   0.80139 1,383
  8k   0.80139 1,522
  7k   0.80139 1,661
  6k   0.80139 1,800
  5k   0.80139 1,939
  4k   0.88153 2,078
  3k   0.97168 2,231
  2k   1.05182 2,399
  1k   1.13197 2,582
  1d   1.22211 2,779
  2d   1.30226 2,990
  3d   1.30226 3,216
  4d   1.30226 3,442
  5d   1.30226 3,667
  6d   1.30226 3,893
  7d   1.30226 4,119
  8d   1.30226 4,345
  9d   1.30226 4,571
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: Strength of Monte-Carlo w/ UCT...

[Andy]

On Sun, Aug 10, 2008 at 3:46 PM, Robert Waite <[EMAIL PROTECTED]>wrote:

> Okay.. so where is the paper that correlates the speed at which MCwUCT
> approaches perfect play with the ability to play a human? They seem
> unrelated as of yet.
>

The closest I've seen are these two studies Don made:

http://cgos.boardspace.net/study/

http://cgos.boardspace.net/study/13/
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] CG'2008 paper: Whole-History Ratings

[Andy]

Remi, you mentioned how the other algorithms predicted well and guessed that
it's because the great majority of games are between experienced players
whose strength is not changing much.  I also feel that the existing KGS
ratings work well for those players already.  So how about focusing on how
the various algorithms perform in the case of improving players.  I think it
would be interesting to simulate game results of various improving players
and show how the different rating algorithms work.

For example:  Suppose a player's true strength is 1500 for some time, and
then he suddenly improves to 2000.  Both before and after he plays a fixed
number of games per day (say 10).  Show a graph of what each rating
algorithm would think his rating is over time.  Many people complain that
the KGS algorithm does not move fast enough for a case like this.

Also the last paragraph of section 4 talks about how the model does not
account for the different ability of new players to change (improve) their
ratings compared to older players.  Could you vary the parameter 'w' based
on the player's current rating?  (Assume players with low ratings are
capable of improving more quickly than strong players).  I don't know enough
about the math to know if this would blow up the computation time or if
that's simply impossible.




On Tue, Apr 8, 2008 at 5:37 PM, Rémi Coulom <[EMAIL PROTECTED]>
wrote:

> Hi,
>
> This is my CG2008 paper, for statisticians:
>
> Whole-History Rating: A Bayesian Rating System for Players of
> Time-Varying Strength
>
> Abstract: Whole-History Rating (WHR) is a new method to estimate the
> time-varying strengths of players involved in paired comparisons. Like
> many variations of the Elo rating system, the whole-history approach is
> based on the dynamic Bradley-Terry model. But, instead of using
> incremental approximations, WHR directly computes the exact maximum a
> posteriori over the whole rating history of all players. This additional
> accuracy comes at a higher computational cost than traditional methods,
> but computation is still fast enough to be easily applied in real time
> to large-scale game servers (a new game is added in less than 0.001
> second). Experiments demonstrate that, in comparison to Elo, Glicko,
> TrueSkill, and decayed-history algorithms, WHR produces better
> predictions.
>
> http://remi.coulom.free.fr/WHR/
>
> Feedback is welcome.
>
> Rémi
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Andy]

But the program isn't stronger than pros, so how can it give better
information about proper komi?

On Feb 11, 2008 6:09 PM, Christoph Birk <[EMAIL PROTECTED]> wrote:

> On Mon, 11 Feb 2008, Don Dailey wrote:
> > I don't bet,  but if I did,  I would bet that it's 7 or 8, and I'm
> > fairly certain that with best play the game would end with 7 extra
> > points for black.
> >
> > I think this was discussed at great length 2 or 3 years ago.
>
> I know ... I brought it up again because of Mogo's success.
> A very (!) strong program should be able to tell us the proper
> komi.
>
> Christoph
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] Transformation between KGS ratings and Elo

[Andy]

See below I created a table that shows the transformation from KGS ratings
to the Elo that CGOS uses.  I set 6k=1800 because I believe that is what GNU
3.7.10 is under both systems.  Does anyone have more data points for bots
that play on both systems?

Also is there an "all times" list for 19x19?  CS-9-17-2CPU is on top at
2297, but I don't think that's the strongest CrazyStone.  I think other strong
19x19 bots are not on the current list either.

To create the table I calculated the probability for a player of rank X to
upset a player of rank X+1.  Then I converted this to the equivalent number
of Elo points for the same upset rate.  Finally I made a running total,
setting 6k to 1800.  The transform could be done using calculus I suppose
but it's been a long time since I did that!  :)


   ELO  set
rank k1 rank  6k=1800
 10k   0.80139 1,244
  9k   0.80139 1,383
  8k   0.80139 1,522
  7k   0.80139 1,661
  6k   0.80139 1,800
  5k   0.80139 1,939
  4k   0.88153 2,078
  3k   0.97168 2,231
  2k   1.05182 2,399
  1k   1.13197 2,582
  1d   1.22211 2,779
  2d   1.30226 2,990
  3d   1.30226 3,216
  4d   1.30226 3,442
  5d   1.30226 3,667
  6d   1.30226 3,893
  7d   1.30226 4,119
  8d   1.30226 4,345
  9d   1.30226 4,571
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Go rating math information

[Andy]

Yes, the math is attempting to model reality.  :)

That's why I also included the EGF data which is based on observed
statistical upset rate.  Of course those ratings are calculated using a
formula which pre-supposes an upset rate.  Round and round.  :)


On Jan 31, 2008 1:26 PM, terry mcintyre <[EMAIL PROTECTED]> wrote:

> From: Andy <[EMAIL PROTECTED]>
>
> Sorry, the KGS formula uses a constant k which is different from the
> K-factor in Elo.
> P(A wins) = 1 / ( 1 + exp(k*(RankB-RankA)) )
>
> This would be equivalent to changing the constant 400 in:
> P(A wins) = 1 / ( 1 + 10^((Ra-Rb)/400)) )
>
> EGF has a similar scheme except of course they use different letters for
> equivalent constants.  So this varying of k is what accounts for the fact
> that upsets are more likely for weak kyu players than for dan players.
>
> I think the varying k factor does not explain ( account for ) the fact
> that kyu players are more likely to win upset victories; I believe it is an
> attempt to  model the inconsistency or variability in kyu-level play.
>
> --
> Looking for last minute shopping deals? Find them fast with Yahoo! 
> Search.<http://us.rd.yahoo.com/evt=51734/*http://tools.search.yahoo.com/newsearch/category.php?category=shopping>
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Go rating math information

[Andy]

A little more information on what CrazyStone's real performance on KGS has
been like during the time the graph depicts:

http://www.gokgs.com/gameArchives.jsp?user=crazystone&year=2007&month=3
http://www.gokgs.com/gameArchives.jsp?user=crazystone&year=2007&month=4
http://www.gokgs.com/gameArchives.jsp?user=crazystone&year=2007&month=11
http://www.gokgs.com/gameArchives.jsp?user=crazystone&year=2007&month=12

In March 2007, CrazyStone starts to play some rated games after a long time
without playing.  It's initially [?] (totally unknown rating).  It starts
with a lucky streak, and gets up to 1k.  Then in April it plays some more,
and the rating settles to weak 2k.  It continues to play many games and
stabilizes there.

Then in November/December 2007, I assume a new version started playing (I
remember reading somewhere it was new version.  Or perhaps just better
hardware).  It quickly settled to a strong 1k.  But in Nov/Dec it played
only 12 rated games, so the sample size is quite small.

Then it quit playing again.  When such a small sample size, it's rating is
very sensitive to "drift".  KGS uses a maximum likelyhood algorithm.  So if
any of CS's 12 opponents gets stronger/weaker, CS will drift with them.  I
assume that is what caused it to move to 1d.

Don asked if he could play some games, stop, and have is rating increase.
The answer is yes if your opponents' ratings increase.  If your opponent's
ratings decrease, you'll go down instead.  Generally for low kyu players,
everyone is improving quickly so there is a strong upward drift.  It's
usually not so bad for the strong kyus or dans since those players don't
generally improve very fast.  However as I said in this case CS has only
played 10 games, so a much smaller pool of opponents is involved.

- Andy

On Jan 31, 2008 12:59 PM, Andy <[EMAIL PROTECTED]> wrote:

> CrazyStone hasn't played since the initial spike to 1k in December.  The
> movement of the chart afterwards is "rating drift".
>
>
> On Jan 31, 2008 12:49 PM, Gian-Carlo Pascutto <[EMAIL PROTECTED]> wrote:
>
> > Don Dailey wrote:
> >
> > > I don't know how David figures 1000 ELO,  but I would expect the
> > > difference to be much larger than that for 19x19 go. I don't
> > believe
> > > they are yet very close to 1 Dan.
> >
> > http://www.gokgs.com/graphPage.jsp?user=CrazyStone
> >
> > You're right. They're closer to 2 Dan.
> >
> > :)
> >
> > --
> > GCP
> > ___
> > computer-go mailing list
> > computer-go@computer-go.org
> > http://www.computer-go.org/mailman/listinfo/computer-go/
> >
>
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Go rating math information

[Andy]

Sorry, the KGS formula uses a constant k which is different from the
K-factor in Elo.
P(A wins) = 1 / ( 1 + exp(k*(RankB-RankA)) )

This would be equivalent to changing the constant 400 in:
P(A wins) = 1 / ( 1 + 10^((Ra-Rb)/400)) )

EGF has a similar scheme except of course they use different letters for
equivalent constants.  So this varying of k is what accounts for the fact
that upsets are more likely for weak kyu players than for dan players.

- Andy


On Jan 31, 2008 12:37 PM, Don Dailey <[EMAIL PROTECTED]> wrote:

> ELO ratings don't have to be absolute, just self consistent.   So if you
> beat someone 7.2% of the time,  that means you are about 440 ELO
> stronger than him.
>
> However, I don't understand what the K-factor has to do with anything.
> scaling it up or down doesn't change anything.  It's common practice to
> make the rating of strong players change more slowly as the result of a
> win or loss but that's not relevant here.
>
> The findings below indicate that differences between dan players is
> greater than the difference between kyu players. So you could not
> assign a fixed ELO per rank but it would have to progressively get
> higher as the players get stronger.
>
> I don't know how David figures 1000 ELO,  but I would expect the
> difference to be much larger than that for 19x19 go. I don't believe
> they are yet very close to 1 Dan.
>
> - Don
>
>
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Go rating math information

[Andy]

CrazyStone hasn't played since the initial spike to 1k in December.  The
movement of the chart afterwards is "rating drift".

On Jan 31, 2008 12:49 PM, Gian-Carlo Pascutto <[EMAIL PROTECTED]> wrote:

> Don Dailey wrote:
>
> > I don't know how David figures 1000 ELO,  but I would expect the
> > difference to be much larger than that for 19x19 go. I don't believe
> > they are yet very close to 1 Dan.
>
> http://www.gokgs.com/graphPage.jsp?user=CrazyStone
>
> You're right. They're closer to 2 Dan.
>
> :)
>
> --
> GCP
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] Go rating math information

[Andy]

There were some questions about the effective ELO difference of two players
3 ranks apart.  Here are some links to information about go rating formulas,
and some statistics:

http://senseis.xmp.net/?KGSRatingMath
http://gemma.ujf.cas.cz/~cieply/GO/gor.html
http://gemma.ujf.cas.cz/~cieply/GO/statev.html

Both KGS and EGF scale the "k" factor according to the player's strength.
For weaker players the probability of an upset is greater.

According to KGS formula:
8k vs 5k: 7.2% chance of upset (~440 Elo)
2d vs 5d: 2.0% chance of upset (~676 Elo)

According to EGF even game statistics:
Generally for weaker kyu players the chance of upset is around 30-40%
(~80-~150 ELO), for stronger players it goes down:
2d vs 5d: 11.5% chance of upset (~350 Elo)
3d vs 6d:  7.8% chance of upset (~432 Elo)
4d vs 7d:  3.3% chance of upset (~590 Elo)
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] cgos 9x9 is stuck?

[Andy Olsen]

The cgos 9x9 server has had two games (66263 and 66262) stuck with no moves
played for about a day now.

I've never used cgosview before, but it's working on the 19x19 server no
problem.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/