Chess2u
Would you like to react to this message? Create an account in a few clicks or log in to continue.

Chess2uLog in

descriptionamazing idea on tourney - the new fixed TC Emptyamazing idea on tourney - the new fixed TC

more_horiz

the top 30% (written in brown color) is the previous writing
from the previous post
but the bottom 70% is the new writing
that is hugely updated and added on
based on the top 30% content

after watching over countless number of engine versus engine matches
for so many years
I have come to the conclusion as follows
this is the opinion and the observation I made
the whole writing regards the TC of a match

***
***

the TC setting is not common
under winboard
I set the TC 6/60
that is 6 moves per 60 seconds
so I thought there should be 6 moves made per minute

but that's not it is working
and the actual TC became much longer

so the 100 or so rounds of super long TC was actually even longer
than I thought and anticipated

1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 5. O-O Nxe4 6. d4 b5 7. Bb3 d5 8. dxe5 Be6 9. Nbd2 Nc5 10. c3 d4 11. Bxe6 Nxe6 12. cxd4 Ncxd4 13. a4 Bb4 14. axb5 Nxf3+ 15. Nxf3 axb5 16. Rxa8 Qxa8

17. Ng5 {+0.43/24} Nxg5 {-0.72/26} 18. Bxg5 {+0.43/21} h6 {-0.80/26}

that's the first 2 moves (4 plies)
H and SF made

but the thing about TC is
when I set 6/60
it is supposed to be 6 moves per minute

but that's not the way it works

after the opening book is used

from the move # 17

one engine has 3 minutes to play only for 2 plies

so SF made only 2 plies for 3 minutes
and H made only 2 plies for 3 minutes

and both engines just spent 6 minutes only to make 2 moves
then from the move number 19

they are given 1 minute for 6 moves

so the actual TC was so much longer than 10 seconds per move
the actual TC on average is something longer than 10 sec per move

*

and the following is the found solution
to solve that above issue
I found a new TC

***

update:

the beta version of winboard 4.70 GUI
is winboard master-20130201
under this beta WB GUI
I tested that 6/60 works fine
both engines are given 1 min to start
and they have 60 sec for 6 moves

but under the complete build of WinBoard-4.7.0 (non beta version)
the TC works differently
I found this amazing thing out
under this non-beta WB GUI
6/60 works differently from the TC under beta WB
like mentioned above

so if I do the super long TC again
I'll use incremental TC
that is 1 min is given to start
then 9 sec are added per move
that should be approx same as 10 sec per move

***
***

the new content update:

(continued from the above post)

after testing for various TCs under the various circumstances
I found out that the above content written in brown
is not that correct
and that brown content contains many errors
so I did a lot of test runs
then I came up with the following corrections
and also improvements upon the current settings

running engine matches
and engine tourneys for so many years
I almost never used
long TC
also never used incremental TC

for many years
I almost always used rapid TC
something like 1 sec per move

***

and using incremental TC
and long TC is like a new area
that should be studied in depth

and I found out the intricate and complex nature
behind long TC and incremental TC

***

update:

for the super long TC tourney
in the rating list
I set the TC 6/60
I thought
that is 6 moves per 60 seconds

but the actual time required was a lot longer than
6 moves per 60 sec

*

then I tried to change the TC
to incremental TC

and the new incremental TC I used was
1 min is given to both engines
then 9 sec per move

but even this incremental TC
is not what I intended

the actual time needed for both engines
was a lot longer than the average of 10 sec per move

***

update:

I found out the complex nature behind
these long TC and incremental TC
and I think I found out how they work

*

the 6/60 TC is tricky to use
(6 moves per 60 sec)
and I can't get the exact TC I want

*

also if I use the incremental TC

1 min is given to both engines
then 9 sec per move

I thought if I use this TC
then on average, the TC should be
approx 10 sec per move

*

but the book depth is 16

I found out that
even when the book move is done
when a move is done from the book
the engines gain increment
so 16 moves from the book * 9 sec per move
= 144 sec

plus the initial 1 min

so, 144 s + 60 s = 204 s

so the engine uses a lot more time
then 10 sec per move
since from the move number 17
they already have 204 sec to begin with

so 1 min is given to both engines
then 9 sec per move

should be changed to
something like
1 min is given
then 7 sec per move
(if we assume the average match length is 60 moves)

then it will be comparable to
10 sec per move

***

for the 6/60 TC
6 moves per 60 sec

book depth is 16

so 6 moves * 2
then 2 min were added
and they start with 1 min
so they already have 3 min

and 6 * 3 = 18
18-16 = 2

they only have 2 moves left
for 3 min

that was the reason why
the engines started with 3min
then they spend total of 6 min
only to make 2 moves (4 plies)

cuz the moves made from the opening book
were counted

***

update:

so, to get the desired TC
I chose another kind of TC

fixed time per move

that is max of 10 sec per move

the only issue with this TC is
the engines usually have time forfeit a lot
when I tested with 1 sec per move

so autoflag is turned off

so the TC will be 10 sec per move
(max of 10 sec per move)
with autoflag off

***

update:

I started the tourney with this new TC
(fixed time per move)

and when many matches are done
(it will take about 1 day
to do 40 rounds
so it will take a few days in total)

I'll post the tourney result
and this entire writing on various TCs
(about the experiments on TCs)
at the same time

*

I think some engines are not time forfeit proof
that is
for some engines
since they don't have the strong functionality
when they have time forfeit
when they go over the time limit
the engines just become malfuctioning

but for the most recent engines
they come with the time forfeit proof function
so they will not be malfuctioning
even when they go over time limit

so the fixed time per move: 10 sec

should be flawless and perfect

also after watching some matches
this fixed time TC looks flawless

***

update:

all the matches will be only done under this new TC
that is fixed time per move

the reason being

I have watched and analyzed countless number of long TC matches
for so many years

and I have come to the conclusion that

some engines think (calculate) for a short amount of time
(= if they have total of 1 hour to think
they only use 40 minutes of 1 hour
they don't even use 20 minutes)

while other engines think for a long time
(if they are given with 1 hour
they use most of that 1 hour)

we can say it is those engines' decision
that they didn't use up all the time allocated and given

maybe the time management is not done well
for those engine
since they only use less than 70% of all the time given

*

I have seen a case like the following case so many times for many years

engine A and B are both 2000 elo strong in rapid TC

but A spends only 70% of the given time
when B uses up almost 100% of the given time

in the rapid TC
like 1 sec per move (60/60)
or 60 moves per 1 min

their elo ratings are same (2000 elo)
since even if A uses a lot less time than B
in a rapid
using a few seconds less doesn't make a big difference in performance
or the searched depths
so the moves found and produced are not that penalized and lowered in quality
even if A uses 30% less time than B does

but when those engines A and B fight
in a long TC
like 2 hours per each (4 hours per round)
using 30% less time can produce a significant different result and the elo rating

so in a long TC
B's elo is 2100
when A's elo is 2000
(but both engines' elo ratings are exactly same in a rapid TC)

*

one example is
houdini 3 uses about 70% of the given time
while SF (the newest SF version) uses almost 100% of the given time

their rapid ratings are almost same

but their long TC ratings are different
(SF is about 50 elo higher)

I thought this is cuz
SF can search very deep (deeper than houdini does)
since SF can search selectively
SF's selectiveness is very high
SF is a fast searcher

but now I think that
SF's elo becomes 50 higher than houdini's elo
cuz SF uses all of 2 hours
(the total match time is about 4 hours for the usual long TC)
while houdini only uses 1 hour and 20 minutes

this is what I observed
also this is what was observed by so many other ppl

*

there are many engines which don't use 100% of the given time
they only use 70% of all the time given
I have seen so many engines like these
maybe they have to change the time management function
but until the time management function is changed
I can do the correction

if I only run engine matches
in the fixed time per move

this kind of problem is resolved

I watched over many matches to confirm this theory

if I set the TC "fixed time per move: 10 (10 or any other numbers)"

when the TC is "fixed time per move:10"
and let's say engine A and B had 100 rounds
and 1 round was 60 moves long on average
so there are 2 plies x 60 moves = 120 plies made per round
and there are 100 rounds

100 rounds * 120 plies = 12,000 = 12k

there are 12k plies (half moves) made during the 100 rounds
all 12,000 plies are made and chosen after using 10 seconds
even if the board position is super simple to solve
the engines only use 10 sec for all those 12,000 situations
all of 12,000 plies were made after using the full 10 sec

so it will be fair
and the elo ratings should be more accurate

*

for SF Nov 11 vs Houdini 3

under the conventional TC
(such as certain number of moves per certain time,
like 60/60, or incremental TC)

the rapid TC elo ratings are almost same

but the long TC ratings are different
60 elo is higher for SF

*

for SF Nov 11 vs Houdini 3

under the "fixed time per move"

the rapid TC elos are almost same

and the prediction is
the long TC elos should be almost same

*

it is just like the university exams
when the exam time is 2 hours
some students finish their exams only in 1 hour
then leave the room
while others use up all of their 2 hours

and those early leavers usually get less scores
we can say
it's their fault to leave early and not using 100% of the given time

but we can't say
the early leavers are less smart than other ppl who stay in the room until the bell rings

let's say there are 2 students
and they are exactly same in intelligence and IQ and the knowledge they possess

if A leaves the room only in 1 hour
and B leaves the room after spending all of the 2 hours

B gets considerably a better score and a higher result

in this case, we can say
it is A's fault that he left the room early

but we can't say
B is smarter and more intelligent than A
cuz they are equal in the abilities and competence

***

and I want to count only those matches
which are done under the same fair condition
that is
all the engines should spend the exact same amount of time
no less, no more

that is the reason why
all the matches will be only done
under this "fixed time per move"

*

let's assume there is a boxing match
players A and B are playing

and when A is knocked down
the referee gives up to 10 sec for A to rise up

whereas when B is knocked down
the referee only gives up to 7 sec for B to get up

I don't want the discrepency between the two to exist
I want both to have the same amount of time
that is 10 sec

*

in conclusion

some engines over 3100 elo have time management issue
they don't use 100% of the given time
they only spend 70~80% of the given time
for these engines
we need something
I will only run the engine matches under "fixed time per move"

it is just a fun fact that
some engines over 3100 elo make simple mistakes and weak moves at the ending
I have seen many times that
some engines with 3100 elo or higher make beginner's mistakes
and for these engines
we need something
we need TBs

when we assign and allocate the exact same amount of time for both black and white
for the whole one match
why not assign the exact same amount of time per move (for all the moves)?
wouldn't it be better and a good idea if all the moves are made for the exact same amount of time?
from this reasoning and notion
all the matches will be run and done
under the fixed time per move TC

descriptionamazing idea on tourney - the new fixed TC EmptyRe: amazing idea on tourney - the new fixed TC

more_horiz


update:

(continued from the above post)

I have watched over
so many long TC matches done by many different ppl
and many different tourneys
between houdini 3 and SF recent (as of Apr. 2013) dev versions
for a long time

for all the matches
SF uses most of the given time
H uses like 70% of the given time

I have seen and read
this observation being voiced out many times by many ppl
for a long time

***

to take one example of the matches

from nTCEC season 1 (about half a year ago)
we can reminisce about H vs SF super final
H won by a slim margin
that was famous

I just chose one match from the tournament
for a visual example



amazing idea on tourney - the new fixed TC Aaa10



amazing idea on tourney - the new fixed TC B10



amazing idea on tourney - the new fixed TC C10



amazing idea on tourney - the new fixed TC D_copy10

descriptionamazing idea on tourney - the new fixed TC EmptyRe: amazing idea on tourney - the new fixed TC

more_horiz

it is a pop quiz
the answer is below
but before reading the answer
we can try to solve this pop quiz question

the question is,

there are 3 kinds of TCs
120/120
fixed time of 1 sec per move
60/60 (60 moves per 60 sec)

which TC will make the tourney shortest?

it is a brain teaser
all of them essentially and technically mean 1 sec per move on average

***
***

60/60
60 moves for 60 sec

for the first 20 moves
it spends 3~2 sec per move

but many matches finish before 60 moves

for the matches that go over 60

from 61st~80th moves
it spends a lot of time
so the length of the match becomes longer

***

fixed TC / 1 sec per move

the total time required is the shorted

***

120/120
120 moves for 120 sec

let's estimate
the average total moves for a match is 50

but with 120/120
for the first 60 moves

it the engine spends a lot of time
like 4~3 seconds per move

***

total amount of time needed

fixed time per move < conventional TCs (like incremental or 60/60)

fixed time of 1 sec per move < 60/60 < 120/120

even if all three of them mean the same on average
at the first glance
it looks like all of them mean 1 sec per move

but fixed time per move needs the least amount of time

by changing from 60/60 to fixed time of 1 sec per move
I could reduce the amount of total time needed by 20%
so before I needed 10 hours for a tourney
I only need 8 hours

amazing idea on tourney - the new fixed TC What_t10

descriptionamazing idea on tourney - the new fixed TC EmptyRe: amazing idea on tourney - the new fixed TC

more_horiz
Permissions in this forum:
You cannot reply to topics in this forum