Dedicated Chess Computers, ram, and retrospective learning

This forum is for general discussions and questions, including Collectors Corner and anything to do with Computer chess.

Moderators: Harvey Williamson, Steve B, Watchman

Forum rules
This textbox is used to restore diagrams posted with the fen tag before the upgrade.
Post Reply
User avatar
appleshampogal
Member
Posts: 126
Joined: Sun Jan 12, 2014 8:53 am

Dedicated Chess Computers, ram, and retrospective learning

Post by appleshampogal »

One thing I've been quite curious about in regards to dedicated chess computers is if in the event the computer may make an inaccuracy or goodness forbid, a blunder which leads down the road to being checkmated by the opponent. Can the computer make sense of this directly after the fact? Now that it has mapped out a variation through playing through... would rewinding the game offer some helpful ram bits to help steer it in (hopefully) a new direction that may be more favorable. That's one thing in my readings that I haven't been able to find much on, but I'm interested to hear what anyone has to say. Especially considering a recent discussion I read about where one's dedicated computer refused to make the same mistake more than once in a certain position. Thoughts appreciated!

Regards,
User avatar
appleshampogal
Member
Posts: 126
Joined: Sun Jan 12, 2014 8:53 am

Post by appleshampogal »

Anyone?
User avatar
Murat
Member
Posts: 189
Joined: Fri Nov 20, 2015 7:49 pm
Location: Canada

Post by Murat »

May I suggest you ask this question in general discussions forum.


All the dedicated chess computer gurus live in there regards ;)
User avatar
Harvey Williamson
Site Admin
Posts: 6079
Joined: Sun Jul 29, 2007 6:57 am
Location: Media City, UK
Contact:

Post by Harvey Williamson »

Murat wrote:May I suggest you ask this question in general discussions forum.


All the dedicated chess computer gurus live in there regards ;)
Good idea - I will move it!
User avatar
Steve B
Site Admin
Posts: 10129
Joined: Sun Jul 29, 2007 10:02 am
Location: New York City USofA
Contact:

Post by Steve B »

Some dedicated computers did have extra Ram for "learning"
in Particular the Fidelity EAG V2-V10 Computers had extra ram to learn from its mistakes

however this feature is not seen in many computers

Here is an example of the EAG "Learning"

[fen]4r2k/q1p2ppp/8/4b2R/3Pn1n1/2P4N/PP4PP/QR5K w - - 0 1[/fen]

when the EAG is set to Level Fixed Depth 4 it will play 1.pxe5
black plays 1.N(g)f2+
the computer will play 2.Nxf2 allowing 2..Nxf2+ leading to mate

if we take back all of the moves and allow the Eag to start over ...
it will now play 1.Rxe5
if we now play 1..Rxe5
the computer now plays 2.pxe5 allowing again 2.N(g)f2+ with the same result

if we take back all of the moves a third time ...
the EAG will now avoid all of the above and play
1.Ng5

the computer "learned " to avoid the previous two bad outcomes

Learning Impaired Regards
Steve
User avatar
appleshampogal
Member
Posts: 126
Joined: Sun Jan 12, 2014 8:53 am

Post by appleshampogal »

Steve B wrote:Some dedicated computers did have extra Ram for "learning"
in Particular the Fidelity EAG V2-V10 Computers had extra ram to learn from its mistakes

however this feature is not seen in many computers

Here is an example of the EAG "Learning"

[fen]4r2k/q1p2ppp/8/4b2R/3Pn1n1/2P4N/PP4PP/QR5K w - - 0 1[/fen]

when the EAG is set to Level Fixed Depth 4 it will play 1.pxe5
black plays 1.N(g)f2+
the computer will play 2.Nxf2 allowing 2..Nxf2+ leading to mate

if we take back all of the moves and allow the Eag to start over ...
it will now play 1.Rxe5
if we now play 1..Rxe5
the computer now plays 2.pxe5 allowing again 2.N(g)f2+ with the same result

if we take back all of the moves a third time ...
the EAG will now avoid all of the above and play
1.Ng5

the computer "learned " to avoid the previous two bad outcomes

Learning Impaired Regards
Steve


That's really interesting. I looked that computer up on wiki to read more about it. How does that work? Is it a specialized line of code to instruct a special purpose operation on what to do with the ram? Or is the certain presence of ram alone herald some kind of learning effect by itself. I'm fascinated. I remember doing a test a little while ago with my challenger where I arrived at a certain result in a particular position, then switched it to "find mate mode" then back to its normal setting and it preferred a different move upon initiating a new computation. It sparked my curiosity if perhaps computers without a learning feature might cache some fractions of ram should they be useful.
User avatar
Cyberchess
Full Member
Posts: 658
Joined: Wed Jan 08, 2014 6:10 pm

Re: Dedicated Chess Computers, ram, and retrospective learni

Post by Cyberchess »

appleshampogal wrote:One thing I've been quite curious about in regards to dedicated chess computers is if in the event the computer may make an inaccuracy or goodness forbid, a blunder which leads down the road to being checkmated by the opponent. Can the computer make sense of this directly after the fact? Now that it has mapped out a variation through playing through... would rewinding the game offer some helpful ram bits to help steer it in (hopefully) a new direction that may be more favorable. That's one thing in my readings that I haven't been able to find much on, but I'm interested to hear what anyone has to say. Especially considering a recent discussion I read about where one's dedicated computer refused to make the same mistake more than once in a certain position. Thoughts appreciated!

Regards,
Greetings Kat:

This is a very interesting topic, and a useful augment to my Repeat Performance Issues With Dedicated Units thread.

The only dedicated unit I’ve ever owned or tested that utilized a learning feature was the Novag Star Diamond. While the contents of the computer’s RAM are volatile – which is to say that they are completely erased whenever the New Game key is pressed or when the unit is powered down, the Star Diamond utilized EPROM memory for its user programmable opening book and for its learning feature. This type of memory is semi permanent as long as working batteries are installed in the unit, but it is also highly sensitive to static buildup which corrupts the data.

The SD had a 10 game learning memory capacity, I believe. I was initially excited about this feature for the reasons cited in my previous post, but soon became disenchanted when I learned of its shortcomings. The unit saves the games in EPROM when its evaluation suddenly drops during a game. Then, in subsequent games, the SD will make an alternative move prior to when its evaluation dropped. In actual practice, however, the SD, not having a clue as to why things went awry, will often change its move after the damage has been done. The alternate move may do little more than delay the undesirable outcome, and the machine will likely expend its available EPROM on different versions of the same lost game.

In summary: while the learning feature is something that needs to be explored further, I feel that without more sophisticated evaluation methods, this feature is somewhat wasted on dedicated units, and the extra EPROM module could have been used to increase the user programmable opening book instead.

~Major Learning Curve~ Regards,
John
User avatar
Steve B
Site Admin
Posts: 10129
Joined: Sun Jul 29, 2007 10:02 am
Location: New York City USofA
Contact:

Post by Steve B »

appleshampogal wrote:
That's really interesting. I looked that computer up on wiki to read more about it. How does that work? Is it a specialized line of code to instruct a special purpose operation on what to do with the ram? Or is the certain presence of ram alone herald some kind of learning effect by itself. I'm fascinated


The computer has ram set aside to store certain positions that resulted in bad or very low eval outcomes
these positions are retained in the computers memory even when turned off and unplugged
the feature can be disabled if the user wants to and it can be reset and wiped clean
AFAIK its basically a question of the hardware(ram) and not really a programming issue although i imagine there must be some code involved

Live And Learn Regards
Steve
User avatar
scandien
Member
Posts: 206
Joined: Mon Sep 12, 2011 1:15 pm
Contact:

Post by scandien »

I remember a winboard engine ( i know this is not a dedicated chess machine) named CILIAN and create by a swiss programmer.

f we trust in the programmer writing ,the chess engine creates his own evaluation table by playing chess. At this time the machine get a rating near 1800-1900.

So it is really possible for a program/machine to learn from experience...

BR

Nicolas
User avatar
Bryan Whitby
Senior Member
Posts: 1001
Joined: Wed Feb 18, 2009 9:57 pm
Location: England

Post by Bryan Whitby »

Maybe Ron can help us here?

Did the Spraklen's use a Monte Carlo Tree Search in their range of Fidelity EAG V2-10 as part of their 'learning' program?

South of France regards
Bryan
Rob Appleby
Member
Posts: 344
Joined: Fri Mar 23, 2018 4:30 am
Location: das Land der eisernen Wurst

Post by Rob Appleby »

scandien wrote:I remember a winboard engine ( i know this is not a dedicated chess machine) named CILIAN and create by a swiss programmer.

f we trust in the programmer writing ,the chess engine creates his own evaluation table by playing chess. At this time the machine get a rating near 1800-1900.

So it is really possible for a program/machine to learn from experience...

BR

Nicolas
Funny reading this just three years later and the strongest engine in the world is entirely self-taught.
nature abhors a vacuum cleaner
Lonniefam
Posts: 1
Joined: Tue Sep 21, 2021 10:16 pm

-

Post by Lonniefam »

I have already looked at LarsenVB. However, I think it is much better to follow a tutorial rather than look at source code with no guidance. I also think that VB is outdated and slow, so not a good language to build a chess engine in.
Post Reply