The New Sudoku Players' Forum

[champagne]

Hi Mike,

Surely a good way to produce seeds different from other players. I never tried it. As you know, I first produce seeds out of the symmetry of given.

[m_b_metcalf]

Surely a good way to produce seeds different from other players. I never tried it. As you know, I first produce seeds out of the symmetry of given.

Well, for i+j > ~5 it doesn't work well. I use both methods. Sometimes both work, sometimes neither.

Mike

[Pat]

In game 295,
i was 0
and j was 1

from game 291
i took my 6.8 "diamond",
-r9c1 ==> "seed" (with 2-answers-or-more)
and this produced my 2 posts (4.0 and 6.8)

it also produced 5.2/3.6/3.4
("more" except i was away;
then again, Patrice had a better 5.2
which did not get into "more")

[enxio27]

Hmmmm. . . I found the -gop option for gsf sudoku that looks (to me) as though it will generate a vicinity search. Can any of you confirm, and also give me a sample command line? I'm not sure of its usage. Here's the pertinent section:

o +/-NX...: a sequence of -N and/or +N operands with option suffixes.
-N: delete all combinations of N clues; +N: add all combinations
of N clues. Puzzles are checked for single solution after each
addition phase. X may be one or more of the following options.
Each occurrence toggles the previous value. Values are inherited
by subsequent operands in the sequence. X may be one or more of:
a assume the last +M clues have already been cleared for @
c use the -Fc format with no filtering, otherwise use the
default -f format, -q constraints, and -e filtering
e do not check off subpuzzles for duplicates
i add implicit (superfluous) candidates, otherwise skip them
n show off/on ops but do not execute
o list one generated puzzle per input puzzle
p the input puzzle is a pattern -- only change original clues
r -nX random {-N+M} operations per input puzzle
@I start at clue index I counting from 1, clearing clues from
from that index on
@ clear the last +M clues and start at the leftmost empty cell
from the cleared clues
{...}xN repeats the bracketed group N times. {...}:N repeats the
bracketed group 0 or more times until the number of clues is <= N.
Some sequences may exhaust memory. {N}* (alternatively {N}!)
repeats {-N+N} until closure (no new puzzles generated). NOTE: a
recalcitrant bug sometimes omits a small number of puzzles in
combined off/on sequences that would otherwise be listed if the
sequence were done separately; if exact results are required (no
omissions) then use separate sequences.

[Pat]

i forget,
may have used it years ago---

try -go{-4+4}p


sudoku -go{-4+4}p -qNF -euniq()*(23==clues)*(1==minimal) -f%%v <seed.TXT >puzzles.TXT

[enxio27]

i forget,
may have used it years ago---

try -go{-4+4}p


sudoku -go{-4+4}p -qNF -euniq()*(23==clues)*(1==minimal) -f%%v <seed.TXT >puzzles.TXT

Thank you, Pat! I'm giving this one a try. The first run appears to be somewhat promising.

I'm still not sure I understand the bracketed numbers ({-4+4}, etc.). Is that the same as the depth? What does it signify?

[m_b_metcalf]

I'm not sure how to make a filter within the search.

Sorry, I can't help you there, as I don't use gsf's codes.

[200e200w]

Hi folks,

This is my strategy that I used during game 303 (and the strategy that I'll use in the future games).

I search for puzzles using a seed grid which is the puzzle posted by dealer or other puzzle with the posted pattern (usually a puzzle from my generation or a puzzle posted by other person).
My searching is done with the (-n,+n) method, using gsf's sudoku generator. I start with n=3, and with the progress of the game I move to n=4.
After generation, I canonicalise all puzzles using champagne's compilation of Gridchecker.
Then, I pre-rate all puzzles using champagne's sk_gsplit program. This does the fast rating and splits the single file into 5 smaller.
After that, I rate all files using champagne's sk_serate program. This rates the puzzles in the SE fashion, but much, much faster.
After that, I select puzzles to rate using Sudoku Explainer. The puzzles that are satisfactory for me are going to submission to the referee.
This ends the process of puzzle generation, canonicalisation and rating.

200e200w

[champagne]

I use for sure a similar process, the succes is widely in the selection of seeds for the next steps.

The main differences with my process are the following

- I generate as seeds for the firt loop puzzles with a symmetry of given
- I have my own +/-n puzzle generator, but this is more a matter of performance

I started in this game and try to improve a +/- n generator with a first "pre-rating" based on the results of the brute force. Splitting if possible the puzzles in three sub files

. a) ER below 4.2
- b) low ED
- c) others
The file a) id rated using a special program keeping only one occurence per ER
The file b) is in stand by for later process
The file c) is processed as in your description and is used for the selection of seeds for the next steps

Note: I never use played puzzles as seed just because I want to see what I can produce alone. I agree that a winning strategy should use them
Note 2: And I did not rate hard puzzles in the lot b) in the game 303 loosing the best ratings.

[m_b_metcalf]

Has anyone ever wondered, when submitting what they hope will be the highest rating in a given game, how likely it is that that is the case? Well, attached is the distribution of the highest rating in each game up to 304. We see that there is little information to be extracted. The most favoured ER is 10.6, where 10.6/10.6/9.4 won five times but 10.6/10.6/10.6 only once. There are some curiosities like 9.2/2.3/2.3, so never give up hope.

Regards,

Mike Metcalf

[dobrichev]

When a game is played by 2 or 3 players and they use/choose very similar strategies, the chance best puzzle to remain undiscovered increases.
btw, can you count the portion of the games where the highest rating is played in a :::more::: submission? Do such precedences exist at all?

[champagne]

When a game is played by 2 or 3 players and they use/choose very similar strategies, the chance best puzzle to remain undiscovered increases.
btw, can you count the portion of the games where the highest rating is played in a :::more::: submission? Do such precedences exist at all?

Hi mladen,

With to-day's strategies applied by several players, I would say that the chances to have a high rating unseen by any of theese players are very low, although I did not catch some games ago the best rating produced by papy999 (but with a fresh code never used before).

I would tell more. All active players use the vicinity as main tool. At the end of the game, if it has been a competitive one, you have a very small number of self-trumping "more". For me this is very similar to the results of the current scan of solutions grids for possible new 17s. The vicinity (and one or two more tools to produce fresh seeds) is so efficient that the target (here possible submissions) is quickly covered especially when the number of clues is relatively low. With more clues, it takes more time just because you have many more puzzles to test

So the strategy issue is mainly to go quicker than other players in the first hours (48 hours,) if you want to win, and then the best strategy is not the same for all pattens. But at the end, to win, it is in most cases compulsory to explore the area where no diamond exists.

[m_b_metcalf]

When a game is played by 2 or 3 players and they use/choose very similar strategies, the chance best puzzle to remain undiscovered increases.
btw, can you count the portion of the games where the highest rating is played in a :::more::: submission? Do such precedences exist at all?

Mladen,
Firstly, the file I posted was missing its lowest entry, 8.9/2.3/2.3, just for the record (so that's the lowest of the highest).

The equivalent file for 'more' entries is attached. I was surprised how many (41) there were, however, in accordance with champagne's remarks, the most recent was in game 257 in September, 2015. Programs have simply got more thorough.

HTH

Mike


P.S. I have not consistently stored all 'more' entries, so this result is a lower bound.

[coloin]

Well in game 086
The "dealer" [me] won because the players didnt find the 11.4 - despite my hints !
Unfortunately I missed a ::: more ::: submission.

[dobrichev]

I fully agree that the vicinity search is proven productive and that all of us have based our strategies on it.

But I persist that there are still too many degrees of freedom how to proceed. Let mention some of them.
- the amount and quality of the initial seed.
- the "leading" rating used for seeding - one of ER, EP, ED.
- using a combination of ER, EP and ED when selecting the seed.
- increasing the depth (distance) on seed exhaustion versus relaxing the rating thresholds.
- using or not using transient non-minimal puzzles.
- perform partial or full pre-rating for each generated puzzle and the thresholds for the partial rating.
- using or not using "downgrade" search (applicable for the intermediate ER and EP ratings).
- playing for score, for best rating, or for rarity/beauty.
- intentionally exploiting the bugs in the official rating engine by submitting particular morph.
- submission strategy (predicting opponent's intention, even to resign).
- dynamics of all the above parameters during the game progress.

I don't believe neither that the top half of the above decisions are equal for all players, nor that they are irrelevant to the chance to reach one of the "absolutely" best rated puzzles for the respective pattern.

Also I am not sure whether the used logarithmic rating is sufficient to explain the high rate of pearls and diamonds in the top submissions. Aren't they biased by the search methods?

On the other hand I highly appreciate Mike's announcements because they are NUMBERS, and the numbers usually are more valuable than the common (non)sense.