Author: Steven Edwards
Date: 11:17:31 04/04/05
Go up one level in this thread
On April 04, 2005 at 13:56:36, Pallav Nawani wrote: >On April 04, 2005 at 03:00:25, Steven Edwards wrote: >>On April 03, 2005 at 23:32:30, Pallav Nawani wrote: >>>On April 03, 2005 at 16:32:35, Steven Edwards wrote: >> >>>>A sample search narrative generated by Symbolic for an opening move selection >>>>can be seen at: >>>> >>>>http://www.geocities.com/chessnotation/Web.html >> >>>Why would you want to do such a thing (ie, create a search narrative) ? >> >>Several reasons: >> >>1. A search narrative provides provides an excellent diagnostic tool for a >>knowledge based program like Symbolic. At the very least, it can show what the >>search was doing immediately prior to a hang or crash. >> >>2. As the program is intended to be psychologically realistic, the natural >>language search narrative can be shown to a sufficiently skilled human player >>for review. The human can assist with improving Symbolic by providing a >>critique of the analysis without having to understand the program's internals. >> >>3. The HTML version of the search narrative, compared with the plain text >>version, can be more easily viewed (e.g., better diagrams and font size >>control). >> >>4. The search narrative proves that the program really is a cognitive searcher >>and not a traditional A/B searcher in disguise. >> >>5. It is possible to use the search narrative in a form of post-search automated >>machine learning by having a stronger program analyze various decision points in >>the narrative and reporting back less than optimal choices. > >The search narrative, as you have it right now, can certainly help in reason 1, >but it dosen't seem to help any for reason 2. For example, see the below text. > >"A total of one second elapsed during the selection. A processor time of zero >seconds was used for the selection. My processor utilization rate is twenty-one >percent. The search tree has one interior node and one leaf node for a total of >two nodes. The maximum variation length is one ply. My selection rate is ten >nodes per processor second, and my mean node processor time is one hundred >milliseconds." > >I doubt if this paragraph will help a human player to tell you anything which >will improve your program's play. But maybe you will add more data to this >later? That pargraph helps me a lot because it tells me how the latest version of the interpreter is performing on a particular platform. Obviously there is much much more to be added, and it's going to take some time. However, the above page does demonstrate that the basic narrative generation method is sound. >You said that your program is a cognitive searcher, does it not use alpha - beta >to search the tree? A/B can be used in a program as it can be used by a human; not applied at uncounted millions of position nodes, but rather at a handful (or several handfuls) of plan evaluation nodes.
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.