Author: Vincent Diepeveen
Date: 13:56:41 08/05/02
Go up one level in this thread
On August 04, 2002 at 15:43:49, Miguel A. Ballicora wrote: >On August 04, 2002 at 15:24:50, Sune Fischer wrote: > >>On August 04, 2002 at 14:46:18, Vincent Diepeveen wrote: >> >>>On August 04, 2002 at 14:06:55, Sune Fischer wrote: >>> >>>will, believe, think, consider. >>> >>>Please proof it. chess is very simple compared to other >>>applications where automatic tuning is supposed to work >>>in the future. >>> >>>So far i have not seen a single decent program that >>>can do better with automatic tuning than without. >> >>Well, maybe you have, do you know how the pros are tuned? >> >>How can it be, that the pros have such a good evaluator, while you being an >>FM(!), can't make something better? >> >>>there is a shitload of freeware programs and volunteers to rewrite >>>them to enable automatic tuning. Please >>>pick a strong program and tune it. I would advice crafty. >>>A small parameter set. Even big advantage for the tuners, >>>but already a good program to start with. >> >>It is a non-trivial exercise to do, and I don't know every character of >>Crafty's code. Besides I would rather spend time on implementing this in my own >>program and get an edge :) >> >>I believe I can make it work, maybe even improve on it. >>It isn't real important to me whether you believe it works or not, I think you >>should follow your ideas and I will follow mine, actually I prefer if you forget >>all about TDLeaf as soon as possible :) >> >>>Finding the best values as a human isn't trivial. It sure isn't >>>for programs. But humans use domain knowledge your tuner doesn't. >> >>KnightCap was too interesting a project not to follow up on, I'm very surprized >>it hasn't been done already. >>To see people write that it doesn't work when a) KnightCap proved it _did_ work, >>and 2) they have not even attempted it themselfs, is very funny to me. >> >>-S. > >5000 parameters is not much when compared to the parameters needed to obtain the >optimal conformation of a protein with a computer. In that case, it is almost >impossible (with the current knowledge) to obtain the right conformation >starting from scratch, but is is very doable when you start from the "optimal" >conformation (determine by physical methods, not by a computer), you change >something and see how the new conformation would look like. Iterations around >the minimum are very fast because all the parameters behave almost linearly or >close enough. When the parameters behave linearly the time to resolve the >problem is O(1) (a linear regresion of n parameters is O(1)). So, even if they >are not perfectly linear, it is way below n log(n). This might not be exactly >like chess but it allows me to make this guess: I believe that learning methods >will be useful to tune all the parameters when a new one has been introduced. >I wonder how many parameters has been throw out in good programs just because >they did not work, but they just needed a general fine tuning of all the rest. > >Regards, >Miguel For the experiments i conducted it didn't parameterize of course all parameters, but i focussed upon a few important parameters. The ones which are tunable from the GUI too. that's about 50 parameters in total. Of course there are 3 start values - a) default values - b) random values - c) all values random What we see in case b and c is such horrors that i don't want to discuss even about it, but in case of a we see that the newlocal maxima it choses never gets even *near* to the local maxima of the default values. I know many who tried TD learning and i only had to look to how knightcap played and to see some writings about what the guy did, to realize what it was doing. The results of knightcap (as interpreted by me), tao, zchess, DIEP, the king and ZZZZZZ and many others are all similar what learning is doing. You toy for a long while and think: "perhaps do it different". If i design a pattern p, then obviously i have in mind it to be positive or negative. In short a bonus or a penalty. Nothing as important as the + and - sign. this is the most important mistake one can make: the + and - sign of a heuristic. Secondly there is the fine tning of the heuristic. obviously when a heuristic is +2.000 it doesn't matter whether a tuner tunes it as +1.852 or +2.425 However when it gets +5.0 then obviously something is WRONG. dead wrong. If a passer bonus which always is positive gets from +1.0 to -0.75, then we talk about a major problem. But the real problem is the lack of domain dependant knowledge. Where each pattern has been typed in by hand, obviously it's a waste of time to use automatic tuning. If i speak for diep i wouldn't know how many parameters i have. It's not important. we talk about tens of thousands probably. It's been years that i counted. Look it's very easy to get bunches of parameters by using arrays to index a heuristic. Where many pro's don't use arrays but a constant value, versus me perhaps an array, that's directly a factor 64 difference in the number of 'adjustable parameters'. I even try to avoid arrays a little, in order to get not a too huge program. It's already not fitting in any PC's L2 cache which ain't funny, though not a big problem (not all parameters get considered each evaluation, you get a kind of logarithmic behaviour in selecting of course which patterns to evaluate so the price of extra parameters is not there simply). Obviously automatic tuning is only useful if it manages to create better local maxima when it runs longer. Does it with your protein research?
This page took 0.01 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.