Collecting labeled positions help wanted

[Ferdy]

Tried to collect positions with tactical move and positional move. This data can be used to classify engine of how tactical or positional it is. This perhaps would be beneficial to engine authors and users.

You can add positions on the following link.
https://docs.google.com/spreadsheets/d/ ... sp=sharing



You may also read the basic guide in Guide sheet.

Later I will write a program to test engines and classify them.

[zullil]

Tried to collect positions with tactical move and positional move. This data can be used to classify engine of how tactical or positional it is. This perhaps would be beneficial to engine authors and users.

You can add positions on the following link.
https://docs.google.com/spreadsheets/d/ ... sp=sharing



You may also read the basic guide in Guide sheet.

Later I will write a program to test engines and classify them.

It might be to your benefit to provide criteria that make a move "tactical" or "positional".

[PK]

I may try to add some positions, but I wonder whether process of finding them can be automated. Say, take two vastly different Rodent personalities (Tal/Ampere for tactical moves and Chloe/Defender for positional ones) and pick positions where they differ.

[Ferdy]

Tried to collect positions with tactical move and positional move. This data can be used to classify engine of how tactical or positional it is. This perhaps would be beneficial to engine authors and users.

You can add positions on the following link.
https://docs.google.com/spreadsheets/d/ ... sp=sharing



You may also read the basic guide in Guide sheet.

Later I will write a program to test engines and classify them.

It might be to your benefit to provide criteria that make a move "tactical" or "positional".

Generally tactical is something that seeks exchanges, or create capture threats whereas positional is something that build good positions or develop pieces first like placing its pieces to restricts opponent's movements.
Try to run 2 engines in analysis mode, one is Lc0 and the other is Stockfish, parse the game, sometimes you will see their preferred move is not the same. Stockfish has the tendency to resolve positions thru tactical means, seeking exchanges and threats at low depths, but with more depth Stockfish may settle for quiet moves. Look at Lc0's prefered move, it usually prefers quiet moves at low depths.

[Ferdy]

I may try to add some positions, but I wonder whether process of finding them can be automated. Say, take two vastly different Rodent personalities (Tal/Ampere for tactical moves and Chloe/Defender for positional ones) and pick positions where they differ.

That is possible indeed and the output position/move can then be inspected visually as verification.

[zullil]

Tried to collect positions with tactical move and positional move. This data can be used to classify engine of how tactical or positional it is. This perhaps would be beneficial to engine authors and users.

You can add positions on the following link.
https://docs.google.com/spreadsheets/d/ ... sp=sharing



You may also read the basic guide in Guide sheet.

Later I will write a program to test engines and classify them.

It might be to your benefit to provide criteria that make a move "tactical" or "positional".

Generally tactical is something that seeks exchanges, or create capture threats whereas positional is something that build good positions or develop pieces first like placing its pieces to restricts opponent's movements.
Try to run 2 engines in analysis mode, one is Lc0 and the other is Stockfish, parse the game, sometimes you will see their preferred move is not the same. Stockfish has the tendency to resolve positions thru tactical means, seeking exchanges and threats at low depths, but with more depth Stockfish may settle for quiet moves. Look at Lc0's prefered move, it usually prefers quiet moves at low depths.

Good luck with your project. I'm not convinced that Lc0 is more "positional" than Stockfish, or even that "tactical" and "positional" have good definitions.

[chrisw]

Tried to collect positions with tactical move and positional move. This data can be used to classify engine of how tactical or positional it is. This perhaps would be beneficial to engine authors and users.

You can add positions on the following link.
https://docs.google.com/spreadsheets/d/ ... sp=sharing



You may also read the basic guide in Guide sheet.

Later I will write a program to test engines and classify them.

try counting proportion of forcing type moves in mainline

for a first approximation:
for epd in epdlist:
mainline = search(epd, stockfish, lc0, whatever)
if mainline exists
count mainline moves, captures, captures lastmovedpiece, checks, promos, whatever
move is tactical if (countwhatever etc / countmoves) > tactics_trigger

[Ferdy]

Tried to collect positions with tactical move and positional move. This data can be used to classify engine of how tactical or positional it is. This perhaps would be beneficial to engine authors and users.

You can add positions on the following link.
https://docs.google.com/spreadsheets/d/ ... sp=sharing



You may also read the basic guide in Guide sheet.

Later I will write a program to test engines and classify them.

try counting proportion of forcing type moves in mainline

for a first approximation:
for epd in epdlist:
mainline = search(epd, stockfish, lc0, whatever)
if mainline exists
count mainline moves, captures, captures lastmovedpiece, checks, promos, whatever
move is tactical if (countwhatever etc / countmoves) > tactics_trigger

Nice idea.

[Ovyron]

Try to run 2 engines in analysis mode, one is Lc0 and the other is Stockfish, parse the game, sometimes you will see their preferred move is not the same.

I can only speak of some laughably low Depth 8 Leela around here, but what I can say is that when they disagree, you can always make them agree by playing their moves to each other, until both agree on the best continuation.

And when Leela has it right and Stockfish doesn't, it's still not clear it was because a positional factor.

It seems more useful names for these moves would be "Evaluation" moves and "Search" moves. The former being the ones where some engines (say, Stockfish) requires really long time to find them, but some other engine can find them really quick because of its evaluation (they "know" the move is best.) While the latter would require long time to search until the tactic is inside the horizon and it is seen and found. The reason these can't be classified as "Positional" and "Tactical" is because at some point new engines will be able to find the moves in the "Search" category really fast, thanks to improvements on evaluation.

So "Search" moves become "Evaluation" moves as chess engine technology advances, something like this can't be used to declare whether a move is Positional or Tactical.