predicting choices of weak humans

[Uri Blass]

I wonder if it is possible to write a chess program that predict choices of humans at different levels.

It seems that for humans not always shorter lines are easier to see.

See the following problem in chesstempo

https://chesstempo.com/chess-problems/58240

The line Rxf6 Nxe4+ Rxc6(call it line A) is only 3 plies to calculate when the line
Rxf6 Nxe4+ King escape Nxf6 (call it line B) is 4 plies to calculate
so I could expect humans who see line B to see also line A that is simpler but it seems that it is not the case and a lot of people blunder at blitz in chess tempo by seeing line B and rejecting Rxf6 and guessing another move to fail the problem(otherwise the problem could get a smaller rating in blitz).

[zullil]

I wonder if it is possible to write a chess program that predict choices of humans at different levels.

Sure. Simply train a neural net using move choices from the appropriate population.

[Ovyron]

Yeah, I had proposed that if imitation is the point, a second network that is an antagonist would help. The second network is trained to differentiate the moves from actual humans from moves from the first network, and the first network gets penalized every time it makes moves that doesn't fool the second network.

You are done when (this is unlike chess strength where you're never done...) the first network produces moves that the second network can't pinpoint as network moves better than chance, because in the way to this, the second network will have become better at telling the difference between a fake move and a human move than any human, and could be used to predict the choices of weak humans.

[EroSennin]

If you don't see it, you just don't see it. Humans don't see anything with 100% accuracy. 64 squares is too many to see at once. Thus we divide the board in to little pieces. Rarely do we go through all the pieces of the puzzle. Some are just left behind as it takes too much time. We aren't machines.

[Ovyron]

Yeah, but people do it systematically, so a machine could be made that mimics human blunders.

Probably the Engine Similarity Tester could be used for this, you just tweak the engine until it plays most similarly to know moves by humans in some rating range.