Saturday, March 31, 2012

Maximizing transfer

.
.
.
.
At the moment I have not the slightest idea how I am going to attack this subject. But the past weeks show that blogging is a very fruitful way to order my thoughts. A lot is clarified already, which helps to add precision in the method of chess improvement. Especially the discovery of the right and the wrong moment to add intelligence is very important.

Question.
Let's see where this will lead to.
An important question that is not really adressed sofar is how to maximize transfer. With "transfer" I mean the the following (in the context of tactics training):
If I exercise problem x in the right way then that will have a positive effect on solving problem y, which I have never seen before, but has something in common with problem x.

Boundaries.
Let's define the extreme boundaries first.
The transfer is 100% if I only have to do 27 problems, one problem per tactical motif, and I will be able to solve every problem at CT without further exercising. I will be able to solve any similar tactical problem in my games as well.
The transfer is 0% if I have to learn all 40,966 problems at CT by heart. It would be useless since it will not transfer to my games either.
In reality the transfer will lie somewhere in between 0% and 100%. It will be clear that maximizing the transfer ratio will cut down efforts and time, and the effect on my games will be maximal too.

What governs transfer.
• Geometrical pattern recognition.
• Analogy.
• Rules
• Conceptualization/Generalisation.
Geometrical pattern recognition.
A main element in transfer is pattern recognition. A pattern that we have learned in problem x is recognized in problem y. A pattern is a rather vague term that can be used for almost anything. I think it might be a good idea to add the sort of pattern we are talking about. In chess most of the time we are talking about geometrical patterns. I referred to the storage of geometrical patterns in memory as "plan A". I don't think it is very useful to use the terms plan A or B any longer. It has served its goal. The basic patterns are the 27 tactical motifs and their combinations.
The patterns are formed by constellations of pieces.

Analogy.
Right now I can't think of an analogy specific to tactics, but analogies are a potential means to transfer knowledge from one problem to another.

Rules.
A rule like "a pinned piece is a bad defender" or "attack the pinned piece"works for certain types of positions.

Conceptualization/generalisation.
With my division of tactical motifs in traps and duplo attacks I tried to catch the essence of the tactical motifs. From there I derived a few questions which can be used when adding intelligence. The questions I'm talking about is about the importance and types of squares. Attacking squares, target squares and focal points. Until now I had focussed on adding intelligence during finding the solution. I will have another look at it when focussing at adding intelligence during the storage phase.
In this specific case of conceptualization we talk about the geometrical pattern recognition of squares, not pieces. The main idea of generalization is to make something applicable between different position which have something (the idea) in common.

My gut feeling tells me that only a few hundred positions (say 400-800) learned in the right way should be enough to master the vast majority of tactics. That's why I focus on higher rated problems and am not worried about the cost of time per problem. But I have no proof for this. Further investigation of transfer is needed.

1. At CT there are meanwhile more than 27 motives.
Of the 27 you are thinking about, 2 are completely useless: "needs more moves" and "different opponent move needed", which says more about the quality of a CT puzzle but does not have to do with the tactic itself. I doubt training these 2 "motives" will lead you anywhere.
Also the motive "exposed king" isnt the best, because it is the most applied tag and is often used is you can give the king a check or if it is not a check, then there are not so many pieces around that king. I would say it is a pretty useless tag/motive, too.

If you go through the list of motives, you might squeeze it down to maybe 10-15 useful motives. True, isnt it?

The more complex and full of motives a puzzle is, the less puzzles you need to cover all motives.
Probably with one complicated puzzle you are able to address 7-8 motives? So you actually might only need to learn 2 or 3 puzzles. Then you covered it all.

Still, I believe you wont improve even if you find these well chosen 3 puzzles and spend 3 hours with them.
I feel it is the wrong approach. It is much better to isolate the motives and see them in many different examples.
Well chosen you might do with 30 examples per useful motive (some motives will need 10 puzzles, others 60)? Then you get the number 300-450 puzzles to be learned. But unfortunately I dont know of such a nice assorted set of puzzles.
And maybe the assumption is wrong and much more than 30 puzzles per motive would be needed?

The low number of 400 (or even 800) looks suspicious to me. No, I dont believe in such a super-set that covers it all. But if I believe into it or not, it is rather a theoretical question, because the super set doesnt exist. Nor is our knowledge equally destributed. Some puzzles are easy for others, but not easy for me. My trouble puzzles are not trouble puzzles for most others. Of the "super set" I would probably be able to solve more than half of them. No intelligence added. It just sounds wrong to me. We need to train what we dont know, and the rest we know we dont need to train. It is wasted time. A super set must be tailored for each of us individually. but still 400-800 puzzles only? I am suspicious.

2. It is possible to learn 40000 CT puzzles. With spaced repetition. It takes a few (~5?) years and i am shure, you will be a great tactician after that training.
I think,.. i will do.

3. @Aox,
It is not impossible to learn 40,000 problems. If I was convinced that it worked I would do it. But if you really need 40,000 problems it means that the transfer rate is very, very low. It is better to investigate how we can improve the transfer rate.

If one does 40,000 problems it means being too lazy to think deep;)

4. @Munich,
see the picture of Sammy Reshevsky in this post. His chess experience at that age, is that the equivalent of 800 problems or of 40,000? Which one is closer?

Somehow he managed to squeeze out the maximum transfer of his modest experience.

Doing less, accomplishing more.

5. I don't think a special problemset is needed. If we manage to squeeze out the maximum possible transfer rate.

6. Definitions:
(To have learned a problem) = (to be able to "solve" { =tell the answer } it quick).

I already can solve many problems quick, so my work is already reduced.
As empirical rabbits experiments and my experiments (and the supermemo algorithm) did show: its (easy) possible to learn 15 new(!) problems a day, thats a workload of less than 50 problems (~less than 50-15 repetitions) a day.
I remember a tactician ( hehehe ;-), who was going to learn 10000 problems but did not succed because he did not use spaced repetition. Just doing 7 circles dont work. i did some experiments on that. If the first repetition is not within 8 days ( thats for me! ), then the attempt is "lost" = there is no speed gain if i do the problem again. See my blog. At the moment, i work on a set with 11000 problems and it looks very promising:

I can see how the learning (=repetition) of problems positivly influence my performance ( at the moment judged by the speed i solve them ) on problems i never saw before (=firsttimer).
There is still some finetunining to do.

If one does 40,000 problems it means being too lazy to think deep;)

Masters "know" good moves without thinking, slow processes like thinking are not that important, the content of the LTM does.

7. I have my son Joey as test person.
He has a CT Blitz rating of 1152.
(I had to persuade him and talk a lot, until he was willing to do CT Blitz mode, and he did 44 puzzles, giving him a RD of 55 or so. So the CT Blitz rating isnt absolutely sure, but the assumption that he is at 1150 wont be to far away from the truth.)

Right at the moment he does two spaced repetition sets:

107 check mate in 2 with CT Blitz range of 800-900

and

111 forks with CT Blitz range of 850-900

There are allready now some interesting things to observe:

Joey is almost as fast as me with these puzzles. He often sees the solution in 8 seconds, while I see it in maybe 5 seconds. In Percent it is a huge difference, but really, if you think about it, 8 seconds isnt so slow.

Also, I started with the check mate in 2 with a range of 800-900, but because he seems to do them so quickly, the 2nd set I allready raised the range to 850-900. If I give him a 3rd set, then I need to raise it further to 900-1000.

My impression is, that he improves currently a lot.

After his approximately 3rd repetition, there is almost no difference in my solution time for a know puzzle and his.

I am not so sure about reshevskys childhood and how many patterns and games he had to play. But I get the impression that the reproduction of new aquired intelligence is much superior to an adult. Joey might be somewhat slower to get anything in his long term memory, but the transfer back into our working memory seem so much better than mine.
If I can keep him motivated, we might see him quickly rise into a different level of chess, and I will need to keep raising the rating range for every new set. And these sets are not very big.

Joey is solving spaced repetition sets on my name/account. Only his blitz rating is under his name, which is Joe_2005 (cause he is born 2005 and is currently 7 years old).
Since CT Blitz rating is usually below the fide estimate, his fide rating might be at 1200-1250. Which would be my estimate too, since I think he is at the weak club level Niveau, which starts at 1200-1250, giving him occasionally the chance to beat a 1400 rated player.

His sets have the name "Joe" or "Joey" in the set name. There you can see in my account in my history his solving times. Pretty fast, hardly slower than our solving time would be.

8. @Aox,
If one does 40,000 problems it means being too lazy to think deep;)

Masters "know" good moves without thinking, slow processes like thinking are not that important, the content of the LTM does.

That is not what I mean. In order to find a method with a higher transfer ratio you have to think deep. Once found, it is a matter of applying it during storage.

9. You both seem to see no reason to maximize transfer?

10. Well, the transfer is really important. I believe the best chance to improve the transfer and to reproduce it in your own games can only be achieved by repetition.

The transfer (recall) of my son is much much superior to mine.
That is why Reshevsky could do with little volume of chess puzzles. But we are adults, it doesnt work for us. We need more volume. Why am I so sure about this?
Well, I improved my CT Blitz rating with my training, I improved my chesscube rating, and I feel it made me stronger, too.
I notice during my games CT puzzles I have learned, so I can clearly see here an effect, caused by the recall from my LTM into my working memory during my games, which is another proof for me. I very much doubt I could have achieved all this with 400-800 asorted puzzles.

11. You both seem to see no reason to maximize transfer?

I see a method wich works, and ehh not that much till now.
I start with the method wich works and try to improve this one.
Look at Empirical rabbits experience about "transfer" here : http://empiricalrabbit.blogspot.de/2011/09/susan-polgar-experiment.html

i never did try to quantify the transfer rate but my experiment with the #11000 Set shows, there is transfer. Soon i will continue analysing my experiences

12. YAWN.

z.

13. I finished a spaced repetition set ranged 1300-1475 CT blitz rating of quite exactly 3000 puzzles (the size number changes though by about +/-1%, probably due to rating changes of the underlying puzzles?). Well "finished" means that I went through and the puzzles scheduled for repetition are getting seldom.

These 3000 puzzles were puzzles I did during my tag-sorted sets, such as "forks", "discoveries" and many more.
It excludes check mate puzzles (non mate puzzles only), it excludes 1000 non-mate puzzles in the range 1300-1475.
(these are puzzles that dont contain any of my trained tags, often these are puzzles with the only tag "hanging pieces" or "exposed king".

So all of these 3000 puzzle (= 75% of the non mate puzzles in the range 1300-1475) I have already seen at least 5 times before during the last 11 months. some of them I have seen more than 20 times, but at least 5 times even those puzzles I could very fast from the very beginning.

But I never put them all together in one set, but only trained them as "fork-Set", "discovery-set" etc.

Guess what: I still do some of these puzzles wrong. And I still dont remember many of them.

I started now a set made of 581 failures (Puzzles I did wrong at least once).
I do this set and wonder: 10% of them I still do wrong. That means I dont even remember the right move!
Most of them I solve correct, but I dont even remember half of them. maybe not even 20% of them I do fully remember. I must have done them a dozen of times at least!
With "remember", I mean remember for sure. When I see them, then I have often a vague feeling that I have seen it before. In this respect I remember many of them. But I fail to remember the whole solution. Often the memory isnt there from the very start but I can most often tell AFTERWARDS that I did have this puzzle before, when I found the solution and click it through.
I intend to do the spaced repetition of my puzzles "I did wrong at least once" again and again (if necessary), until I know them inside out, until I remember the solution literally, until I can do almost all of them within 6 seconds.

If there is a super-set like Tempo like to postulate it, then it must be my set of 581 failed puzzles. I want to really know them.

For the 3000-set I finished I looked up: How many of them I solved within 9 seconds the last time I have seen the puzzle (option: "use most recent time")?
I solved 1780/3000 within 9 seconds. That means that I needed 10 seconds or more for 1220/3000. This is somehow scary, and I am not content with this at all. I expected much more than just "60% I can do fast".

I intend to concentrate even more on my failed puzzles, and drop the "slow solved" puzzles for later.

14. @Munich
You can see at my last post in my blog that the repetitionrate of CT's SRS is to low for "complicated" ( ~ = failed) problems for me.
I need to repeat them more often. I guess/estimate 10% more repetitions can do at this set. As more "comlicated" a problem as more often i need to repeat it.

In Supermemo-speech: the ease factor is to high

I think, if you solve a problem slower than before, then the space between the repetitions of this problem have to get smaler (not bigger), but richard dont allow the growth of space to become lower than 1.3. So even if i get slower at solving a problem, i have to wait for a repetition 30% longer (minimum!). But if you get slower than you already start forgetting!

I did send Richard a message about it. He intend to do some changes in the (near?/far?) future. I hope many tacticians share my interest at SRS and give a feedback to Richard too ;-)