Author: Georg v. Zimmermann
Date: 15:12:56 11/21/03
Go up one level in this thread
Hi Dustin, I think you raise some very interesting questions there. On November 21, 2003 at 16:21:35, Dustin Moore wrote: >One basic assumption seems to be that engines assume the opposing >player will play the "best" move that the engine would play were it >playing for the other side. > >Is there anything that can be gained by breaking this assumption? >Let's say an engine is on a search extension or a quiescence search >out to some deep ply can it skip calculating some of deepest plys for >the opposing machine if it knows that the other machine doesn't >search as deep normally? One 'problem' I see with this idea is that when I am already searching deeper regularely, I do not need any more fancy tricks. I will win anyway. > >Or a more obvious method of exploiting this. Let's say one engine, >engine A is playing engine B. Wouldn't A have an edge if it talked >with another engine B' that was running the same software as B so >it could figure out how the other engine would respond? Interesting as well. The problem here is time. Say every engine has 3 minutes per move. Engine A now thinks for 2 seconds, then decides it might want to play 1.e4. It talks with B's brother : what would you do if I'd play 1.e4 ? Unfortunately B's brother would have to reply : just give me 3 minutes, then I will know. But A does not have the time to wait 3 minutes ... > > >Another question, most the algorthims based on the mini-max >(alpha-beta/MUD/whatever) are oriented twords finding the one branch with >alternating global maxes and mins out to some depth. > >What if there was an additional search (in addition to all the standard >ones that a good engine like crafty does) that was oriented twords going >very deep but not hoping to find a global max. Let's say I have two >CPU's working the problem. One CPU is a standard crafty that runs out to >14ply and the other that intelligently tries single branches out to >20-ply in the hopes of finding a single branch of the tree that is better >than what the main crafty's current evaluation. The principle is, if >both competitors are drawn at a 14-ply or whatever calculation, it only >takes a single "supernatural move" to change the course of the game. This >move doesn't have to be the best move that a 20ply computer would make, >just better than what the 14 ply computer has at the time. It would sort of >be a modification of the alpha beta algorithm. This second "CPU" would have to, to get a possible variation the first CPU could "believe" in, be no more selective for the defending color than the first CPU. It could be more selective for the other color. This idea I actually know that it not only can work, but also it has been done.If I remember correctly, the author of one of the strongest suicide chess programs (you know where you have to capture all the pieces) told me that he is always doing a very short proof-number (I think it was) search to try to find forcing wins (comparable to mate), and if that fails, do a normal alpha-beta. I think you see the similarity. In normal chess I fear that the benefit of just using the 2CPU's normally is greater .. but maybe not, how about you try it out ? :-) >Instead of cutting off the >search when all of the branches have been proven to have scores less than >the first or second branch's max, cut off the search and move another >ply deeper when a branch has been found that is better than the current >14 ply crafty calculation. > >I guess I'm just trying to figure out ways to get deeper searches at the >expense of complete searches. I hope the second idea doesn't come off as >just the same thing as a search extension whicn I don't think it is. A >search extension happens when favorable conditions are found for extending >the search. I think you are right there. > My idea is to ALWAYS be searching random deep branches in the >hopes that some game-turning discovery is found. > >Lastly, I'm sure that both of my ideas arise from basic >mis-understandings about the way things work. I'm just interested in the >comments given the amazing knowledge of this group.
This page took 0 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.