What is the largest difference between the "best" move from a given position as evaluated by a chess engine and the worst move regularly played by strong players, maximized over all opening positions regularly reached? (See King's Indian Defense as an example.)
We can also obtain ground truth on the other end, by providing positions known to be drawn according to a large tablebase, but only providing a small (or no) tablebase to the engine. What is the spread of these evaluations?
These margins of error can be used to create a stochastic opponent which (weighted) randomly picks among top moves within the margin (using MultiPV).
No comments :
Post a Comment