To see if Romi is learning you can start Romi from the command line and type pr 1 enter. Romi will print out for example, "Record 1 sib 487 chd 2 fs 12 ts 28 some other stuff score 17 w 283 l 264, d 191" which is 1.e2e4. Then type pr 487 in this example to get the record for the first sibling move. In this example it is d2d4, has a score of 19 w 23 l 22 d 13. anyway it is possible to examine the learn file starting with "pr 1".
The resigning problem is new as far as I know. It is an issue between the compiler I use and Windows 7/vista and windows 8/10. The resigning problem does not happen on w10 or as far as I know w8.
Thanks for the info! Was that already documented somewhere?
BTW exists a possibility to see the content as whole in an ascii export?
(something like PG dump for Black/White into ascii files)
Another question, I guess the format hasn't changed since you introduced
it and we can still experiment with the old and big learn files e.g. Marc Lacrosse and Pavel Hase made?
I also have Win7 here and will take a look at the resign problem, which is
yet unknown to me - I was also almost away from computerchess from 2011-2016 ;-)
I have downloaded Romichess for the first time to see what all the hype is about with the reinforcement learning.
using RomiChess64P3n.exe
created as a winboard 2 engine in arena
learn.dat file autocreated on creation of the engine
100 game tournament setup with a computer opponent
tournament is ongoing but learn.dat is still 274KB and not growing
What did I do wrong?
So far Romichess is 0/33
Thanks!
To see if Romi is learning you can start Romi from the command line and type pr 1 enter. Romi will print out for example, "Record 1 sib 487 chd 2 fs 12 ts 28 some other stuff score 17 w 283 l 264, d 191" which is 1.e2e4. Then type pr 487 in this example to get the record for the first sibling move. In this example it is d2d4, has a score of 19 w 23 l 22 d 13. anyway it is possible to examine the learn file starting with "pr 1".
The resigning problem is new as far as I know. It is an issue between the compiler I use and Windows 7/vista and windows 8/10. The resigning problem does not happen on w10 or as far as I know w8.
The abortive resign problem also happens on my Win 8.1 system, albeit less frequently. Would it be possible to compile Romi locally from source? If so, what would I need to use to compile it?
By the way, I can verify that Romi's learning file updates normally after each game (using Arena 3.5). I created a very high adjudication threshold to give Romi a chance to resign in a lost position before GUI adjudication kicks in.
Guenther wrote:
I also have Win7 here and will take a look at the resign problem, which is
yet unknown to me - I was also almost away from computerchess from 2011-2016 ;-)
Guenther
Yes it happens an awful lot here. Obviously it comes from the learn file.
May be a wrong value is read (used a learn file from 2007) which triggers resign?
All moves before the resign come from the learn file...
Why not just disable auto resign (until the cause is found)? It is an obsolete feature nowadays, because the GUIs have advanced resign features anyway.
I have downloaded Romichess for the first time to see what all the hype is about with the reinforcement learning.
using RomiChess64P3n.exe
created as a winboard 2 engine in arena
learn.dat file autocreated on creation of the engine
100 game tournament setup with a computer opponent
tournament is ongoing but learn.dat is still 274KB and not growing
What did I do wrong?
So far Romichess is 0/33
Thanks!
To see if Romi is learning you can start Romi from the command line and type pr 1 enter. Romi will print out for example, "Record 1 sib 487 chd 2 fs 12 ts 28 some other stuff score 17 w 283 l 264, d 191" which is 1.e2e4. Then type pr 487 in this example to get the record for the first sibling move. In this example it is d2d4, has a score of 19 w 23 l 22 d 13. anyway it is possible to examine the learn file starting with "pr 1".
The resigning problem is new as far as I know. It is an issue between the compiler I use and Windows 7/vista and windows 8/10. The resigning problem does not happen on w10 or as far as I know w8.
The abortive resign problem also happens on my Win 8.1 system, albeit less frequently. Would it be possible to compile Romi locally from source? If so, what would I need to use to compile it?
By the way, I can verify that Romi's learning file updates normally after each game (using Arena 3.5). I created a very high adjudication threshold to give Romi a chance to resign in a lost position before GUI adjudication kicks in.
CL
I compiled with MSVS 2005. I used nameless structs in the code which seem to be an MS extension only. If you can find MSVC 6 it will compile 32 bits. I really like nameless structs because amazing relationships can be defined and then eloquently accessed.
Now that I know it is a learn file problem thanks to Guenther maybe I'll be able to find it.
If you are on a sidewalk and the covid goes beep beep
Just step aside or you might have a bit of heat
Covid covid runs through the town all day
Can the people ever change their ways
Sherwin the covid's after you
Sherwin if it catches you you're through
To see if Romi is learning you can start Romi from the command line and type pr 1 enter. Romi will print out for example, "Record 1 sib 487 chd 2 fs 12 ts 28 some other stuff score 17 w 283 l 264, d 191" which is 1.e2e4. Then type pr 487 in this example to get the record for the first sibling move. In this example it is d2d4, has a score of 19 w 23 l 22 d 13. anyway it is possible to examine the learn file starting with "pr 1".
The resigning problem is new as far as I know. It is an issue between the compiler I use and Windows 7/vista and windows 8/10. The resigning problem does not happen on w10 or as far as I know w8.
Thanks for the info! Was that already documented somewhere?
BTW exists a possibility to see the content as whole in an ascii export?
(something like PG dump for Black/White into ascii files)
Another question, I guess the format hasn't changed since you introduced
it and we can still experiment with the old and big learn files e.g. Marc Lacrosse and Pavel Hase made?
I also have Win7 here and will take a look at the resign problem, which is
yet unknown to me - I was also almost away from computerchess from 2011-2016
Guenther
Hi Guenther, I think that I saw this but was too tired to answer. Last three days I spent most of my time in bed trying to keep my strength up. Sorry for the late reply.
Was the pr command documented? Yes, before I got sick way back when and did not have the energy except to release the executables of P3k. However, it can be understood from io.c rather easily.
There is no way in Romi to dump the information. However, the structure of the learn file can be easily understood so one can create a standalone program to mine whatever is required.
Yes, the old learn files are valid, zero changes since then.
Glad to see you back! It would be easy to disable the outdated resign feature for the book. However, I'd also want to figure out why there is apparently wrong value getting into the learn file. And why it never happens in Windows 10. If I can not figure it out I will just not retrieve a record with a faulty value in it.
Sorry again for the late reply.
If you are on a sidewalk and the covid goes beep beep
Just step aside or you might have a bit of heat
Covid covid runs through the town all day
Can the people ever change their ways
Sherwin the covid's after you
Sherwin if it catches you you're through
Thank you Graham for making RomiChess available to everyone!
If you are on a sidewalk and the covid goes beep beep
Just step aside or you might have a bit of heat
Covid covid runs through the town all day
Can the people ever change their ways
Sherwin the covid's after you
Sherwin if it catches you you're through
Now it'll be harsh times for Romi as she'll begin training against the muses, how many games until she manages to get a win?
That is a good question!
If you are on a sidewalk and the covid goes beep beep
Just step aside or you might have a bit of heat
Covid covid runs through the town all day
Can the people ever change their ways
Sherwin the covid's after you
Sherwin if it catches you you're through
What I tried was training Romi "bottom-top" instead of "top-bottom".
That is, instead of having Romi play a stronger opponent, she plays a weaker opponent (that she is always able to defeat), and then I increase the strength of the opponent until they're at the same level (i.e. increase the opponent's Depth until Romi draws or loses a game, then Romi trains against the opponent at this depth until she can beat them, and I increase its depth again.)
I was playing Romi at 10 seconds/move against Houdini Typhoon (Tactical=1, Contempt=10) and Romi won all her games up until Depth=7, at Depth=8 the game was drawn.
However, I'm unable to merge this PGN to the learn file, even though I hadn't problems adding 1.2 million games.
Romi just hangs when I try to merge the PGN, can you check it out? Are you able to merge it to a learn file? Am I doing something wrong?
It would be easy to disable the outdated resign feature for the book. However, I'd also want to figure out why there is apparently wrong value getting into the learn file. And why it never happens in Windows 10. If I can not figure it out I will just not retrieve a record with a faulty value in it.
