Reply to Re: I gave up to the Computers National Olimpiad
If you don't have an account, just leave the password field blank.
A genetic algorithm that takes 11 days to calculate one vector. Or did you improve already?
We improved it
It was 11 days because the method that had to calculate the value of a square on the board relied on a try/catch construction which is nice but too slow when you call it millions of times. Basically the code needed the catching part too often but we won't do that in the future anymore (in time critical situations anyway). For those unfamiliar with try/catch, you basically "try" the code and "catch" an exception when it goes wrong (like divide by 0 or other exceptions). It kinda makes sense because (but this I don't know, it's just what I figured) when "trying" some code, you need a backup of the variables just before you start trying the code because when it goes wrong, you need to revert to those again.
We rewrote it and now the time has reduced to 1 - 3 secs for each game (and we need to play 18750 games). Because of this reduction, it takes something like 10 hours to evolve that vector with a search tree depth of 2 so it's likely that we're going to increase the depth of the search tree to get a better vector in the end. Now we're 'only' doing one of our own moves and one countermove by the opponent but of course, the deeper we can look in the tree, the better the decision will be
We improved it
It was 11 days because the method that had to calculate the value of a square on the board relied on a try/catch construction which is nice but too slow when you call it millions of times. Basically the code needed the catching part too often but we won't do that in the future anymore (in time critical situations anyway). For those unfamiliar with try/catch, you basically "try" the code and "catch" an exception when it goes wrong (like divide by 0 or other exceptions). It kinda makes sense because (but this I don't know, it's just what I figured) when "trying" some code, you need a backup of the variables just before you start trying the code because when it goes wrong, you need to revert to those again.We rewrote it and now the time has reduced to 1 - 3 secs for each game (and we need to play 18750 games). Because of this reduction, it takes something like 10 hours to evolve that vector with a search tree depth of 2 so it's likely that we're going to increase the depth of the search tree to get a better vector in the end. Now we're 'only' doing one of our own moves and one countermove by the opponent but of course, the deeper we can look in the tree, the better the decision will be
