Because that will fail to detect a program that halts in X+1 time. The problem isn’t to detect if a program that halts halts, the problem is to generally create an algorithm that will guarantee that the analyzed program will always halt given an infinite time running on an infinite computer.
But you could also do a mean time analysis on specific tasks and have it cut off at a standard deviation or two (90-98% of task times covered), and have a checkbox or something for when the user expects longer times.
You could probably even make this adaptive, with a cutoff at 2x the standard time, and updating the median estimate after each run.
I have always wondered why the answer to the halting problem isn't: "If no output has been returned in X time, BREAK, restart program from beginning."
what if it needed just one more second to complete?
Damn Vogons.
Because that will fail to detect a program that halts in X+1 time. The problem isn’t to detect if a program that halts halts, the problem is to generally create an algorithm that will guarantee that the analyzed program will always halt given an infinite time running on an infinite computer.
Yes, that is the theoretical rational, but in use cases, if users always restart the program after X minutes, practical you can just hard reboot
But you could also do a mean time analysis on specific tasks and have it cut off at a standard deviation or two (90-98% of task times covered), and have a checkbox or something for when the user expects longer times.
You could probably even make this adaptive, with a cutoff at 2x the standard time, and updating the median estimate after each run.