TalkChess.com

[Álvaro Begué]

I'll describe an idea through an example.

Say you have to play 3 moves in the next 20 minutes and then you have 30 minutes for the next 30 moves, then 30 minutes for the rest of the game. You can decide how much time to use based on either:
* 3 moves in 20 minutes
* 33 moves in 50 minutes
* full game in 80 minutes

Because I want the behavior to converge to sudden death if the number of moves is very large, I cap the number of moves at 30, which is what I use for sudden death (of course you can use a different number or heuristics for this):
* 3 moves in 20 minutes => 400 seconds/move
* 33 moves in 50 minutes => 100 seconds/move (after limiting the number of moves to 30)
* full game in 80 minutes => 160 seconds/move (after limiting the number of moves to 30)

Now take the minimum of those numbers: 100 seconds/move.

That's it. This mechanism doesn't suffer from the problem Jon described, and it also doesn't suffer from pathological behavior if the time control is 1,000,000 moves in 30 minutes (spending 1 minute is more reasonable than 1.8 milliseconds).

In Ruy-López we also have a hard time limit to make sure we don't lose on time if the search gets stuck in a deep subtree, and that hard limit is set by looking only at the next time control.