Tuesday, February 5, 2013

My personal (non) progress in tactics



Some History

I did start solving tactical puzzles because its necessary ( but not suficient ) to be strong in tactics to become a better player. No improvement in tactics - > no ( big ) improvement in chess. The trainingsituation in tactics is exellent, there are tactic-books, tactic-software, tactic-servers. The logical first step seemed to be, to improve in tactics a few hundred points. But that is not easy. After more than 100 000 puzzles at several tacticsservers i did start solving puzzles at chesstempo.
















This plot is a recalculated rating based only on problems i saw the first time ( so i dont measure my memory but my tactical skill ) and based on the rating of these problems of today ( to bypass ratingdrift issues ). The calculation is done with a ratingsystem which is "as close as possible" related to the blitz-ratingssystem of Chesstempo. The original ratings of chesstempo does have to many issues with things like "duplicate reward reduction"  to see whats realy going on.

Discussion of the plot

The raise in rating at the beginning is "not real". The ratingsystem of CT is different to the ratingsystem of other server. I did get used to the system of ct, the board, the colours, the pieces, the method how to move the pieces and the special ratingssystem ( for ex. Ct is punishing thinkingtime "after the first move" extra hard ). I did plateau for a while at a level <~1900. The improvement above ~1900 seems to be the result of being in the saltmines = "High speed, high volume, low rated problems - training". After a long vacation from the saltmines i did start making salt again ( in a little modified way ) and my rating is raising again.

Fritz Board Vision Training

Attack training: >35
Defence training: ~30
I will start Check training soon.

I was wondering why/how i did improve in these "Board Vision" exercises.
At the beginning i made many errors, i did remember typical errors ( the type of position where they happen ), i was looking for these errors during my training and... they did vanish, they did get automated. Easy thoughts, done over and over again, get automated. I recognise the position and see ( = remember ) what to do.

Vision = memory triggert by recognition (of chunks).

11 comments:

Munich said...

It is difficult to part the non-duplicates (=first-timers) form the duplicates. However, this is neccessary to get a really accurate reading of the rating based on non-duplicates.

It is not enough to simply filter out any duplicates and look at the remaining puzzles (= fide estimate). You need to recalculate to where the rating based on first-timers would rise. Then, from this elevated level, you need to recalculate how much points you earned (or lost) from this elevated level. While your rating rises, the reward on each puzzle would become smaller.

This is an interative process. You recalculate from puzzle to puzzle. The RD will be a problem, too, however I dont think it is a big problem.
Simply assume an RD of 35 in the formular. It would not make much sense to me to EVER lower the RD anyway.
Does the uncertainty of our rating really become high again? No, certainly not!
A high RD makes sense when we dont know if a person has 1200 in rating or 2000. So wie start with a rating of 1500 and a high RD. But once it reached an RD of 35 (for instance at a rating of 1897), why should the RD rise ever again? Huge jumps in ratings are not to be expected, even if we do get better after a hard training.
To measure progress, you can savely assume an RD of 35 and dont worry about it. It doesnt have to match a theoretical "what would the blitz rating read if only first-timers were included and we let RD rise to unrealistic highs again".
So dont aim for that. Simply aim for the better "what would the blitz rating read if only first-timers were included with a fix RD of 35 after you solved 1000 puzzles".

Munich said...

It is difficult to part the non-duplicates (=first-timers) form the duplicates. However, this is neccessary to get a really accurate reading of the rating based on non-duplicates.

It is not enough to simply filter out any duplicates and look at the remaining puzzles (= fide estimate). You need to recalculate to where the rating based on first-timers would rise. Then, from this elevated level, you need to recalculate how much points you earned (or lost) from this elevated level. While your rating rises, the reward on each puzzle would become smaller.

This is an interative process. You recalculate from puzzle to puzzle. The RD will be a problem, too, however I dont think it is a big problem.
Simply assume an RD of 35 in the formular. It would not make much sense to me to EVER lower the RD anyway.
Does the uncertainty of our rating really become high again? No, certainly not!
A high RD makes sense when we dont know if a person has 1200 in rating or 2000. So wie start with a rating of 1500 and a high RD. But once it reached an RD of 35 (for instance at a rating of 1897), why should the RD rise ever again? Huge jumps in ratings are not to be expected, even if we do get better after a hard training.
To measure progress, you can savely assume an RD of 35 and dont worry about it. It doesnt have to match a theoretical "what would the blitz rating read if only first-timers were included and we let RD rise to unrealistic highs again".
So dont aim for that. Simply aim for the better "what would the blitz rating read if only first-timers were included with a fix RD of 35 after you solved 1000 puzzles".

AoxomoxoA wondering said...

I did recalculate my rating based only on first-timers at this plot. There is no need for any iteration. RD is fixed in my calculation ( and lower than 35 ) that is mathematicaly "the same" as in the ELO-system with a low k. This rating is more than good enough to compare my tactical performance at different times. At this moment i am close to my ATH.

Munich said...

So then --> you seem to have it done correctly. I am just confused:
"There is no need for any iteration."

So when you have a rating of 1900 and you did a first-timer with a blitz rating of 1850 correct and you earned +2.5 points for solving that correctly, then you need to continue with a a rating of 1902.5, right?
In between you get duplicates and you fail them. Your blitz rating falls to 1850. You get again a 1850 rated first timer puzzle, and because of your lower blitz rating you earn 3.0 points.
However, this is incorrect for your "first-timers-only" view.
Correct would be to take your 1902.5 rating and calculate what reward you deserve from this level. It might only be ~2.5 again.
So your rating would rise from 1902.5 to 1902.5+2.5 = 1905 rating points.
Incorrect would be to calculate 1902.5+3.0 = 1905.5

So you dont only need to adjust the rating, but also the reward for solving the puzzle correctly. "Iterative" is probably the wrong word for it, but what I mean is, that you have to recalculate the rating for each step, keeping the RD the same, recalculating the reward based on you calculated "first-timer-rating".
The recalculating the reward on the recalculated rating sound like an 1-step iteration to me, but I am not mathematically sound enough to know when we call an iteration an iteration. I mean that one number is dependend on an other number, whereas both numbers change by influencing each other.

AoxomoxoA wondering said...

I have a rating of 1900 and earn 2.5 points, then i have 1902.5.
Then i get some duplicates, i ignore that, its like looking a nice movie. Now i get this 1850 and i get a score which would give me 3.0 points if i would have a rating of 1850. But my rating is still 1902.5 so i dont get 3,0 but say 2.4 so my new rating will be 1904.9

i calculate a score and then the new rating.
i dont use the new rating calculated by richard.

if i fail then score = 0
ELSE
if time_used>average_time then
score=average_time/TIME_USED
else
Score=1;

http://chesstempo.com/user-guide/en/tacticRatingSystem.html#blitzRating

And now with Elo:

Expecteted Score Ea=1/(1+10^(Rating_problem - My_Rating)/400 )

( if Rating_problem=my_rating then the expected score = 50%, if the rating of the problem is extreme high then the expected score is 1/(1+a_lot)= small )

and now:

New_Rating=Old_Rating+k*(Score-Expected_score)

if i score higher than expected then i gain a lot of points.

http://en.wikipedia.org/wiki/Elo_rating_system#Mathematical_details

Iteration ::= a procedure in which repetition of a sequence of operations yields results >>successively closer<< to a desired result

I just calculate the new rating in one step, i dont get closer and closer to a result by repeating over and over ( usually ad infinitum!) some claculations.

Munich said...

well done. Very good.

Last question: over what time do we look at the above graph (x-Axis) aproximately?

Temposchlucker said...

You are interested in ATH's, which represent the maximal statistical anomalies. I'm more interested in averages, or a Bezier curve, since those help to identify trends.

AoxomoxoA wondering said...

X = No of Blitzattempts ( max ~ 45 000 )

Time ~ 3+ Years,the left half ~2+ Years, the right half of the graph is ~ 1 year

AoxomoxoA wondering said...

The rating is already a "sliding average" like a 20 ( 10? ) days line..

New_Rating=Old_Rating+k*(Score-Expected_score)

as smaler the k as more "average". I use a smaller k than ct in its rating , this rating is already more "trendy" than CT's rating.

I am interested to see: "when did my trend change" then i may(??) find +,- parameters of my training.

Munich said...

Since the right half is only ~1 year, the graph is a bit distorted, and hence the upward trend looks less step (upwards) than in the left half.
So your trend looks going indeed more upward than your graph pictures it!

--> your upwards trend is still intact I'd dare to say.
The interesting thing: even though you adjusted your training within these 3 years - no matter what you do, it goes (very slowly) upwards!

Temposchlucker said...

@Munich,
If you are right, I must be biased. I don't see progress in the right half. That's why I wanted a trendline. Just to be sure.