Author: Robert Henry Durrett
Date: 18:30:35 06/14/02
Go up one level in this thread
On June 14, 2002 at 12:18:56, Dann Corbit wrote:
>On June 14, 2002 at 12:09:34, Robert Henry Durrett wrote:
>>_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
>>
>>Posted by Dann Corbit on June 14, 2002 at 11:18:06:
>>
>>"An Honest Way to Cheat":
>>
>>Implement learning. All computer programs should do this anyway. If they don't
>>it is a defect and the program deserves to be punished.
>>
>>_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
>>
>>A FEW QUESTIONS:
>>
>>(1) How much benefit is **possible** in chess engines by incorporating
>>"learning"? Also, how to measure that benefit?
>
>At infinity, given infinite storage, a learning program will play perfect chess.
>
>>(2) Are there poor, ineffective, ways to do it and good, effective, ways?
>
>Sure.
>
>>(3) What would be the **best** way to incorporate learning in chess engines?
>
>Nobody knows. Here is what has been tried:
>1. TD-Lambda learning is used to adjust the weights for internal calculations
>during play. Examples are BACE, ExChess, KnightCap.
>2. Book learning will remember bad lines, and continually gather statistics
>about book lines during play.
>3. Position learning will recall information about specific board positions.
>
>Are there others? These are the only ones I know about.
Please forgive me for this really bad "cheap shot," but the huge number of
bulletins and literature on this topic would seem to suggest that more would be
known. [:)?]
>
>>(4) Perhaps it would be necessary to distinguish between "knowledge based"
>>learning and other kinds of learning?
>
>What does this mean? Learning is nothing more than an accumulation of
>knowledge. This accumulated knowledge is used to make better and better
>decisions.
What I had in mind maybe can best be explained by discussing anologous
situations in human life and human chess. Starting with relatively trivial
examples:
(a) It's easier to develop one's vocabulary and grammar in a foreign language
if the learner already knows something about the language. The learner utilizes
[perhaps subconsciously] knowledge already available to spot new information.
This is a simple case of "knowledge based learning."
(b) The highly touted "Scientific Method" could be viewed as a process whereby
scientists build upon prior knowledge to discover new knowledge. Without the
prior knowledge, progress is much slower.
(c) Staying with the "Scientific Method" for a minute, consider also the
problem of learning how to learn better/faster. Experienced scientists utilize
their "lessons learned" about more effective ways to pursue new knowledge.
Experience can be a great teacher in any human pursuit, perhaps due to the great
flexibility of the human mind.
(d) Moving now to chess: It would be difficult to learn new ways to restrict
an opponent's game if the chessplayer have never heard of [or independently
thought of] the concept of space advantage. The search for specific techniques
or lines aimed at taking advantage of a space advantage is driven by knowledge
of the space advantage concept. It is that knowledge which motivates the
[human's] search for specific ways to build up the bind. The search would never
begin at all without that knowledge.
(e) I admit that I am not intimately knowledgeable about the "guts" of chess
engines. [You knew that already.] However, it seems to me that maturation of
"learning" in chess engines is not going to be something as trivial as adjusting
a few parameters. Gut feel only.
(f) Etceteras . . .
Of those you listed, I find "position learning" to be most interesting. One
must find ways to identify and recognize "similarities" between known positions
[or "position fragments"] {stored in the computer somehow?} and the position
being examined? Innovation, it would seem, would surely be required to keep
this from blowing up into an unmanageable process.
>
>>(5) Is this already discussed sufficiently in other ICD CCC threads or in the
>>literature somewhere? If so, where?
>
>Scattered all over the place.
I knew that. But the search engine doesn't search recent bulletins. All those
"old" bulletins are probably obsolete anyway. :)
Bob D.
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.