Author: Dave Gomboc
Date: 17:56:01 06/21/99
Go up one level in this thread
On June 21, 1999 at 20:39:05, Dann Corbit wrote: >On June 21, 1999 at 20:27:02, Paul Richards wrote: > >>On June 21, 1999 at 14:02:36, Dann Corbit wrote: >> >>>I have gone over all the pathways calculated in C.A.P. for the position in >>>question (the one I thought was a blunder at first blush). It turns out that by >>>examining the results at subsequent plies you can find better alternatives. >>>This points to (I think) the necessity of doing the "connect the dots" >>>experiment. It also has generated a fascinating hypothesis of mine, which is >>>"The Chess Analysis Project will reinvent all the great chess openings by post >>>processing." >>>And a corollary: >>>"The Chess Analysis Project will invent new openings superior to any in >>>existance." >> >>Had to change the title. :) By the first hypothesis do you mean you will >>simply rediscover what is accepted opening theory? >That is pretty much unavoidable, since all positions from classical openings are >in the database. When you examine single entries, they don't always follow the >classical lines, but when you string them together they tend to. > >>Or that you will >>determine which lines are soundest? >That also. In addition, the lines will be depened, and some will find clear >refutations. > >>How exactly does this sort of solving work anyway? I mean, I can see it >>in general terms, but I'm not sure of the details. Suppose you're at a >>certain position, and you see that you have five moves from that point >>in your database. They have centipawn values of +50, -75, +123, +64, and >>+10 for the side on move. What is the value of your current position? >>An average? If you assume that your opponent makes the best move then I >>suppose the value must be the least, or -75 in this case. If you apply >>this to each point in the database, you should develop long variations >>based on this "least" (best move for the other guy) value at each point. >>Is that about right? And does it matter where you start, or can you >>locate the "ends" of each variation? Then you have to consider how to >>add new data into a "solved" database. Or make a correction. Suppose >>you do a longer analysis and a new point has the least value. I suppose each >>point can have a "raw" original data value and a "solved" value, both of which >>could be updated as needed. Interesting >>stuff anyway. >How it works is pretty simple. Consider some position x. From x, we have 20 >choices, of which 7 are in the database. Currently, for x we have a computer >estimated ce of y. Now, we examine those 7 subsequent positions we already >have. Considering now as the opponent, we let the opponent "do their worst" and >create an estimated ce and pv from that (leaving the originals intact). We >continue with this process, reversing roles from ply to ply, all the way to the >end of the database. Then we repeat the process (perhaps 1000 times). The >reason we repeat it is that each position in the database gets updated in pass >one. Now, looking at our revamped estimates of the worth of a position, what >was our best choice last time may no longer be the best choice. We may also >find that choices extrapolated to be the best have never been attempted. So we >might calculate them and add these in for subsequent runs. Using techniques >like this, we can actually form strategic plans, since we could be looking >dozens of full moves into the future. You already have the leaf values computed, so the proposed algorithm is inefficient. Simply do a depth-first search (with cycle detection) and backup scores in a minimax fashion. Now it takes one pass through the items. Dave
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.