Training with RomiChess in console mode

[BrendanJNorman]

I set Romi to 2 ply and I won against Romi easily. Then Romi played my moves back against me in the next game.

Okay, so Romi could work by learning from a PGN Database from the site where one plays, then Romi would play the winning moves of the people from there, and, at least until a novelty is played, one would be playing against the strongest humans from the site.

I play at lichess, here's their games...

https://database.lichess.org/

Whoa! The games just from the month of November are 3.6GB in size! This would be a big project, so I wonder if someone else that plays at Lichess would be willing to feed the PGNs to Romi and give this concept a try

I suspect a large percentage of those games are of low-quality. I'd rather not fill up Romi's learning file with junk, if it were me. Just the opposite, in fact.

Depends on the goals.

Could copy a new version of Romi into a brand new folder (so as not to pollute your main Romi installation/learn file), rename it "Romi 2000" and populate the learning file only with LiChess games of players 2000-2100.

This might be an interesting way to create a "human-like" opponent who is Advanced Club Player level.

In fact, I might perform this experiment myself.

[Tobber]

I set Romi to 2 ply and I won against Romi easily. Then Romi played my moves back against me in the next game.

Okay, so Romi could work by learning from a PGN Database from the site where one plays, then Romi would play the winning moves of the people from there, and, at least until a novelty is played, one would be playing against the strongest humans from the site.

I play at lichess, here's their games...

https://database.lichess.org/

Whoa! The games just from the month of November are 3.6GB in size! This would be a big project, so I wonder if someone else that plays at Lichess would be willing to feed the PGNs to Romi and give this concept a try

I suspect a large percentage of those games are of low-quality. I'd rather not fill up Romi's learning file with junk, if it were me. Just the opposite, in fact.

Depends on the goals.

Could copy a new version of Romi into a brand new folder (so as not to pollute your main Romi installation/learn file), rename it "Romi 2000" and populate the learning file only with LiChess games of players 2000-2100.

This might be an interesting way to create a "human-like" opponent who is Advanced Club Player level.

In fact, I might perform this experiment myself.

I'm not sure it will make any difference either way. I merged the 2.2 million games pgn into the learn file and also merged last season TCEC into it.
Played 300 games against Spike 1.4. The result was 3% wins, exactly what could be expected with the +600 Elo difference from CCRL ratings.
Romi did build a reasonable opening book from the merged games but is clueless as soon it's out of the book. There are to many possible moves later into the games so the possibility that Romi will find a good move in the learn file is small.

Against a human player with a limited repertoire of openings things may be different. For engine to engine play this learning file method seems to be of limited value. I have started a new tournament where Spike don't have any book but so far Romi seems not be able to take advantage of it.

/John

[BrendanJNorman]

I set Romi to 2 ply and I won against Romi easily. Then Romi played my moves back against me in the next game.

Okay, so Romi could work by learning from a PGN Database from the site where one plays, then Romi would play the winning moves of the people from there, and, at least until a novelty is played, one would be playing against the strongest humans from the site.

I play at lichess, here's their games...

https://database.lichess.org/

Whoa! The games just from the month of November are 3.6GB in size! This would be a big project, so I wonder if someone else that plays at Lichess would be willing to feed the PGNs to Romi and give this concept a try

I suspect a large percentage of those games are of low-quality. I'd rather not fill up Romi's learning file with junk, if it were me. Just the opposite, in fact.

Depends on the goals.

Could copy a new version of Romi into a brand new folder (so as not to pollute your main Romi installation/learn file), rename it "Romi 2000" and populate the learning file only with LiChess games of players 2000-2100.

This might be an interesting way to create a "human-like" opponent who is Advanced Club Player level.

In fact, I might perform this experiment myself.

I'm not sure it will make any difference either way. I merged the 2.2 million games pgn into the learn file and also merged last season TCEC into it.
Played 300 games against Spike 1.4. The result was 3% wins, exactly what could be expected with the +600 Elo difference from CCRL ratings.
Romi did build a reasonable opening book from the merged games but is clueless as soon it's out of the book. There are to many possible moves later into the games so the possibility that Romi will find a good move in the learn file is small.

Against a human player with a limited repertoire of openings things may be different. For engine to engine play this learning file method seems to be of limited value. I have started a new tournament where Spike don't have any book but so far Romi seems not be able to take advantage of it.

/John

I've just downloaded a million games (July 2014) from LiChess and chose only those between 2000-2100 Elo players.

This turned out to be only about 5000 games, and I made a learn file from it.

I played some blitz with it and it seemed pretty fun.

Definitely a decent partner for training games.

[pgn][Event "Computer chess game"]
[Site "BRENDANNORMD8A2"]
[Date "2018.01.02"]
[Round "?"]
[White "RomiChess p3k 2000"]
[Black "brendannorman"]
[Result "*"]
[BlackElo "2400"]
[ECO "D02"]
[Opening "Queen's Pawn"]
[Time "19:01:19"]
[Variation "London"]
[WhiteElo "2000"]
[TimeControl "60+1"]
[Termination "unterminated"]
[PlyCount "87"]
[WhiteType "program"]
[BlackType "human"]

1. d4 Nf6 2. Nf3 d5 3. Bf4 c6 4. g3 {+0.50/13 3} Bg4 5. Bg2 {+0.79/14 3}
Bxf3 6. Bxf3 {+0.73/14 3} e6 7. c3 {+0.62/14 3} Bd6 8. Bxd6 {+0.43/14 2}
Qxd6 9. Qb3 {+0.65/14 3} Qc7 10. Nd2 {+0.41/14 3} O-O 11. c4 {+0.20/13 3}
Nbd7 12. O-O {+0.19/13 3} b6 13. Rad1 {+0.44/12 2} Rfe8 14. Nb1 {+0.49/13
2} Rac8 15. Nc3 {+0.42/14 2} Qd6 16. Bg2 {+0.45/13 1} Qb8 17. Qa4 {+0.56/13
2} h6 18. Rfe1 {+0.63/12 2} Nf8 19. f4 {+0.76/12 1} g6 20. h3 {+1.05/13 2}
N8d7 21. Rf1 {+0.95/13 2} Kg7 22. Rde1 {+1.04/12 2} Qb7 23. a3 {+0.76/12 2}
a6 24. Qb3 {+0.64/12 2} b5 25. c5 {+0.55/13 1} a5 26. Ra1 {+0.50/12 1} Qa7
27. a4 {+0.77/14 2} b4 28. Nd1 {+0.51/14 2} Qa6 29. Bf3 {+0.67/13 1} Ne4
30. Bxe4 {+0.80/15 1} dxe4 31. Re1 {+0.42/15 1} f5 32. Ne3 {+0.51/14 1} Nf6
33. Nc4 {+0.37/14 1} Nd5 34. Nd6 {+1.02/14 1} Kf6 35. e3 {+1.82/14 1} Rg8
36. Nxc8 {+1.78/14 1} Qxc8 37. Rac1 {+1.78/13 1} g5 38. Kh2 {+1.82/13 1}
Qe8 39. Qc2 {+1.78/13 1} Qh5 40. Qf2 {+1.60/13 1} g4 41. h4 {+1.36/14 1}
Qf7 42. Qf1 {+1.36/14 1} Qb7 43. Kg1 {+1.32/13 1} h5 44. b3 {+1.28/14 1} * [/pgn]

Final position is a positional draw, so I ended the game there instead of shuffling for 100 moves.

[Tobber]

I set Romi to 2 ply and I won against Romi easily. Then Romi played my moves back against me in the next game.

Okay, so Romi could work by learning from a PGN Database from the site where one plays, then Romi would play the winning moves of the people from there, and, at least until a novelty is played, one would be playing against the strongest humans from the site.

I play at lichess, here's their games...

https://database.lichess.org/

Whoa! The games just from the month of November are 3.6GB in size! This would be a big project, so I wonder if someone else that plays at Lichess would be willing to feed the PGNs to Romi and give this concept a try

I suspect a large percentage of those games are of low-quality. I'd rather not fill up Romi's learning file with junk, if it were me. Just the opposite, in fact.

Depends on the goals.

Could copy a new version of Romi into a brand new folder (so as not to pollute your main Romi installation/learn file), rename it "Romi 2000" and populate the learning file only with LiChess games of players 2000-2100.

This might be an interesting way to create a "human-like" opponent who is Advanced Club Player level.

In fact, I might perform this experiment myself.

I'm not sure it will make any difference either way. I merged the 2.2 million games pgn into the learn file and also merged last season TCEC into it.
Played 300 games against Spike 1.4. The result was 3% wins, exactly what could be expected with the +600 Elo difference from CCRL ratings.
Romi did build a reasonable opening book from the merged games but is clueless as soon it's out of the book. There are to many possible moves later into the games so the possibility that Romi will find a good move in the learn file is small.

Against a human player with a limited repertoire of openings things may be different. For engine to engine play this learning file method seems to be of limited value. I have started a new tournament where Spike don't have any book but so far Romi seems not be able to take advantage of it.

/John

I've just downloaded a million games (July 2014) from LiChess and chose only those between 2000-2100 Elo players.

This turned out to be only about 5000 games, and I made a learn file from it.

I played some blitz with it and it seemed pretty fun.

Definitely a decent partner for training games.

[pgn][Event "Computer chess game"]
[Site "BRENDANNORMD8A2"]
[Date "2018.01.02"]
[Round "?"]
[White "RomiChess p3k 2000"]
[Black "brendannorman"]
[Result "*"]
[BlackElo "2400"]
[ECO "D02"]
[Opening "Queen's Pawn"]
[Time "19:01:19"]
[Variation "London"]
[WhiteElo "2000"]
[TimeControl "60+1"]
[Termination "unterminated"]
[PlyCount "87"]
[WhiteType "program"]
[BlackType "human"]

1. d4 Nf6 2. Nf3 d5 3. Bf4 c6 4. g3 {+0.50/13 3} Bg4 5. Bg2 {+0.79/14 3}
Bxf3 6. Bxf3 {+0.73/14 3} e6 7. c3 {+0.62/14 3} Bd6 8. Bxd6 {+0.43/14 2}
Qxd6 9. Qb3 {+0.65/14 3} Qc7 10. Nd2 {+0.41/14 3} O-O 11. c4 {+0.20/13 3}
Nbd7 12. O-O {+0.19/13 3} b6 13. Rad1 {+0.44/12 2} Rfe8 14. Nb1 {+0.49/13
2} Rac8 15. Nc3 {+0.42/14 2} Qd6 16. Bg2 {+0.45/13 1} Qb8 17. Qa4 {+0.56/13
2} h6 18. Rfe1 {+0.63/12 2} Nf8 19. f4 {+0.76/12 1} g6 20. h3 {+1.05/13 2}
N8d7 21. Rf1 {+0.95/13 2} Kg7 22. Rde1 {+1.04/12 2} Qb7 23. a3 {+0.76/12 2}
a6 24. Qb3 {+0.64/12 2} b5 25. c5 {+0.55/13 1} a5 26. Ra1 {+0.50/12 1} Qa7
27. a4 {+0.77/14 2} b4 28. Nd1 {+0.51/14 2} Qa6 29. Bf3 {+0.67/13 1} Ne4
30. Bxe4 {+0.80/15 1} dxe4 31. Re1 {+0.42/15 1} f5 32. Ne3 {+0.51/14 1} Nf6
33. Nc4 {+0.37/14 1} Nd5 34. Nd6 {+1.02/14 1} Kf6 35. e3 {+1.82/14 1} Rg8
36. Nxc8 {+1.78/14 1} Qxc8 37. Rac1 {+1.78/13 1} g5 38. Kh2 {+1.82/13 1}
Qe8 39. Qc2 {+1.78/13 1} Qh5 40. Qf2 {+1.60/13 1} g4 41. h4 {+1.36/14 1}
Qf7 42. Qf1 {+1.36/14 1} Qb7 43. Kg1 {+1.32/13 1} h5 44. b3 {+1.28/14 1} * [/pgn]

Final position is a positional draw, so I ended the game there instead of shuffling for 100 moves.

The big question is of course if you can see any difference with or without the learn file from Lichess.

/John

[BrendanJNorman]

I set Romi to 2 ply and I won against Romi easily. Then Romi played my moves back against me in the next game.

Okay, so Romi could work by learning from a PGN Database from the site where one plays, then Romi would play the winning moves of the people from there, and, at least until a novelty is played, one would be playing against the strongest humans from the site.

I play at lichess, here's their games...

https://database.lichess.org/

Whoa! The games just from the month of November are 3.6GB in size! This would be a big project, so I wonder if someone else that plays at Lichess would be willing to feed the PGNs to Romi and give this concept a try

I suspect a large percentage of those games are of low-quality. I'd rather not fill up Romi's learning file with junk, if it were me. Just the opposite, in fact.

Depends on the goals.

Could copy a new version of Romi into a brand new folder (so as not to pollute your main Romi installation/learn file), rename it "Romi 2000" and populate the learning file only with LiChess games of players 2000-2100.

This might be an interesting way to create a "human-like" opponent who is Advanced Club Player level.

In fact, I might perform this experiment myself.

I'm not sure it will make any difference either way. I merged the 2.2 million games pgn into the learn file and also merged last season TCEC into it.
Played 300 games against Spike 1.4. The result was 3% wins, exactly what could be expected with the +600 Elo difference from CCRL ratings.
Romi did build a reasonable opening book from the merged games but is clueless as soon it's out of the book. There are to many possible moves later into the games so the possibility that Romi will find a good move in the learn file is small.

Against a human player with a limited repertoire of openings things may be different. For engine to engine play this learning file method seems to be of limited value. I have started a new tournament where Spike don't have any book but so far Romi seems not be able to take advantage of it.

/John

I've just downloaded a million games (July 2014) from LiChess and chose only those between 2000-2100 Elo players.

This turned out to be only about 5000 games, and I made a learn file from it.

I played some blitz with it and it seemed pretty fun.

Definitely a decent partner for training games.

[pgn][Event "Computer chess game"]
[Site "BRENDANNORMD8A2"]
[Date "2018.01.02"]
[Round "?"]
[White "RomiChess p3k 2000"]
[Black "brendannorman"]
[Result "*"]
[BlackElo "2400"]
[ECO "D02"]
[Opening "Queen's Pawn"]
[Time "19:01:19"]
[Variation "London"]
[WhiteElo "2000"]
[TimeControl "60+1"]
[Termination "unterminated"]
[PlyCount "87"]
[WhiteType "program"]
[BlackType "human"]

1. d4 Nf6 2. Nf3 d5 3. Bf4 c6 4. g3 {+0.50/13 3} Bg4 5. Bg2 {+0.79/14 3}
Bxf3 6. Bxf3 {+0.73/14 3} e6 7. c3 {+0.62/14 3} Bd6 8. Bxd6 {+0.43/14 2}
Qxd6 9. Qb3 {+0.65/14 3} Qc7 10. Nd2 {+0.41/14 3} O-O 11. c4 {+0.20/13 3}
Nbd7 12. O-O {+0.19/13 3} b6 13. Rad1 {+0.44/12 2} Rfe8 14. Nb1 {+0.49/13
2} Rac8 15. Nc3 {+0.42/14 2} Qd6 16. Bg2 {+0.45/13 1} Qb8 17. Qa4 {+0.56/13
2} h6 18. Rfe1 {+0.63/12 2} Nf8 19. f4 {+0.76/12 1} g6 20. h3 {+1.05/13 2}
N8d7 21. Rf1 {+0.95/13 2} Kg7 22. Rde1 {+1.04/12 2} Qb7 23. a3 {+0.76/12 2}
a6 24. Qb3 {+0.64/12 2} b5 25. c5 {+0.55/13 1} a5 26. Ra1 {+0.50/12 1} Qa7
27. a4 {+0.77/14 2} b4 28. Nd1 {+0.51/14 2} Qa6 29. Bf3 {+0.67/13 1} Ne4
30. Bxe4 {+0.80/15 1} dxe4 31. Re1 {+0.42/15 1} f5 32. Ne3 {+0.51/14 1} Nf6
33. Nc4 {+0.37/14 1} Nd5 34. Nd6 {+1.02/14 1} Kf6 35. e3 {+1.82/14 1} Rg8
36. Nxc8 {+1.78/14 1} Qxc8 37. Rac1 {+1.78/13 1} g5 38. Kh2 {+1.82/13 1}
Qe8 39. Qc2 {+1.78/13 1} Qh5 40. Qf2 {+1.60/13 1} g4 41. h4 {+1.36/14 1}
Qf7 42. Qf1 {+1.36/14 1} Qb7 43. Kg1 {+1.32/13 1} h5 44. b3 {+1.28/14 1} * [/pgn]

Final position is a positional draw, so I ended the game there instead of shuffling for 100 moves.

The big question is of course if you can see any difference with or without the learn file from Lichess.

/John

Sure I can.

In the opening, I feel like I'm against a human, not an infallible engine's opening book.

This alone makes it worthwhile for training purposes.

Just my opinion, but lots of ideas come to mind.

Gonna put this guy against some weaker Rodent personalities and see how they play together.

[Tobber]

100 more games Romi - Spike, Spike without book. This time Romi managed to get 9% which equals 150 Elo better. Of course the sample is to small but we have some indication. If it's the learning or Spike's missing book which have most impact I don't know.

/John

[Michael Sherwin]

I set Romi to 2 ply and I won against Romi easily. Then Romi played my moves back against me in the next game.

Okay, so Romi could work by learning from a PGN Database from the site where one plays, then Romi would play the winning moves of the people from there, and, at least until a novelty is played, one would be playing against the strongest humans from the site.

I play at lichess, here's their games...

https://database.lichess.org/

Whoa! The games just from the month of November are 3.6GB in size! This would be a big project, so I wonder if someone else that plays at Lichess would be willing to feed the PGNs to Romi and give this concept a try

I suspect a large percentage of those games are of low-quality. I'd rather not fill up Romi's learning file with junk, if it were me. Just the opposite, in fact.

Depends on the goals.

Could copy a new version of Romi into a brand new folder (so as not to pollute your main Romi installation/learn file), rename it "Romi 2000" and populate the learning file only with LiChess games of players 2000-2100.

This might be an interesting way to create a "human-like" opponent who is Advanced Club Player level.

In fact, I might perform this experiment myself.

I'm not sure it will make any difference either way. I merged the 2.2 million games pgn into the learn file and also merged last season TCEC into it.
Played 300 games against Spike 1.4. The result was 3% wins, exactly what could be expected with the +600 Elo difference from CCRL ratings.
Romi did build a reasonable opening book from the merged games but is clueless as soon it's out of the book. There are to many possible moves later into the games so the possibility that Romi will find a good move in the learn file is small.

Against a human player with a limited repertoire of openings things may be different. For engine to engine play this learning file method seems to be of limited value. I have started a new tournament where Spike don't have any book but so far Romi seems not be able to take advantage of it.

/John

The authors of Spile once posted that Spike learns internally and that learning cannot be turned off. That is why Romi cannot make progress against Spike. It learns also. But that was a long time ago so I could be remembering wrong.

[Guenther]

The authors of Spile once posted that Spike learns internally and that learning cannot be turned off. That is why Romi cannot make progress against Spike. It learns also. But that was a long time ago so I could be remembering wrong.

Yes you remember wrong. Posititional learning always could be turned off in the config file.
May be you mixed it up with the internal book, but that can be turned off too.

[Ovyron]

That is why Romi cannot make progress against Spike. It learns also.

But they use different approaches, and Spike would never play the opponents' moves against them. It would be interesting to see who gets a win first against stronger opposition. Probably Romi could have the edge since she can learn from millions of games at a time, while Spike cannot.

[Michael Sherwin]

That is why Romi cannot make progress against Spike. It learns also.

But they use different approaches, and Spike would never play the opponents' moves against them. It would be interesting to see who gets a win first against stronger opposition. Probably Romi could have the edge since she can learn from millions of games at a time, while Spike cannot.

Yes, maybe. But at the moment I'm kind of burnt out. What I wanted for Romi is to be a good engine for humans to train against and except for a couple of diehard fans that looks pretty much like a flop.