What if Tiffany doesn’t play optimally, or plays in accordance with a different definition of ‘optimal’? (Suppose its maker played every game of tick tac toe, and recorded their response to every situation, then hardcoded those values.)
If Tiffany’s performance is good enough, then Tiffany is still best described as optimising for tic tac toe performance, because:
Tic tac toe performance has high predictive power as an hypothesis for Tiffany’s utility function.
Tic tac toe performance has relatively low complexity when compared to other hypotheses with comparable predictive power.
This changes if Tiffany’s performance is not sufficiently high (in which case there may be some other low complexity objective function that Tiffany is best described as optimising).
What if Tiffany doesn’t play optimally, or plays in accordance with a different definition of ‘optimal’? (Suppose its maker played every game of tick tac toe, and recorded their response to every situation, then hardcoded those values.)
If Tiffany’s performance is good enough, then Tiffany is still best described as optimising for tic tac toe performance, because:
Tic tac toe performance has high predictive power as an hypothesis for Tiffany’s utility function.
Tic tac toe performance has relatively low complexity when compared to other hypotheses with comparable predictive power.
This changes if Tiffany’s performance is not sufficiently high (in which case there may be some other low complexity objective function that Tiffany is best described as optimising).