Gameplay post-mortem

[erakis17]

I decided to write a boggle solver to analyze my gameplay (5x5). Don't worry! It only works by parsing the post-game page, which I save to html. Cheating isn't the slightest bit tempting, and I'm not playing for trophies anyway.

I've previously generated wordlists in order of frequency, for my haphazard study. That assumed typical boggle dice letter frequency, which seems to be approximately true for this game, but I'll be able to check it now.

Some interesting early findings: I don't actually get palindromes as automatically as I thought I did. For example, I got SEER 74% of the time, and REES only 46%, though I have them in my mind as palindromes! On the other hand, I seem to have SESE/ESES and SATI/ITAS down fine.

Words I technically know, but never apparently use:
TAIS, TRES, ERST, UTES, etc.

Others that I didn't actually know, though I'm sure I tried learning them at some point:
REOS, ONIE, ONST, etc.

There are all sorts of study possibilities. Obviously, start looking for those high-frequency words I never get. But a lot of those occur in extremely rich boards anyway, where I have my fingers full just typing. More interesting might be looking for relatively common words among low-scoring boards.

Once I'm more sure of letter frequencies, I'll Monte Carlo a gazillion boards and start building word trees to study. My words list is just SOWPODS, so it doesn't find the very long words. I suppose I could randomly pick one of the long-word lists. But without knowing exactly how wordtwist salts the board with long words, my Monte Carlo will be deficient in them anyway.

There are implications in this: e.g. as many of you probably know, you do well to study even shorter words that show up as pieces of long words,
RATIO
LATI
ZATI
HABLE
and the like. So probably I'll pick a long-word list and make up a separate algorithm to Monte Carlo boards contingent on having a long word.

Anyway, I have a lot of possibilities here, so if there are ideas for generating study lists people have, feel free to comment!

[I play every board, so these boards should be representative for someone who plays that way]
Out of ~330 games analyzed, here are the most frequent words. Format is:

Word .. appearances ..% of times I found the word

TIES ... 56 ... 53.6%
ESNE ... 53 ... 54.7%
NEST ... 52 ... 63.5%
SITE ... 51 ... 74.5%
STIE ... 51 ... 56.9%
SEER ... 50 ... 74.0%
REES ... 50 ... 46.0%
SESE ... 49 ... 75.5%
ESES ... 49 ... 73.5%
SATI ... 49 ... 53.1%
ITAS ... 49 ... 53.1%
SENS ... 49 ... 73.5%
SEEN ... 48 ... 58.3%
SETA ... 48 ... 64.6%
ATES ... 48 ... 39.6%
NESS ... 47 ... 74.5%
ERES ... 47 ... 29.8%
SERE ... 47 ... 80.9%
SIRE ... 46 ... 63.0%
SETS ... 45 ... 40.0%
TAIN ... 45 ... 68.9%
SEAT ... 45 ... 46.7%
TAES ... 45 ... 46.7%
REIS ... 44 ... 25.0%
NITE ... 44 ... 59.1%
SERS ... 44 ... 40.9%
EAST ... 44 ... 70.5%
SIEN ... 43 ... 30.2%
TAIS ... 43 ... 0.0%
TEAS ... 43 ... 48.8%
SESS ... 43 ... 46.5%
RETE ... 43 ... 65.1%
TEER ... 42 ... 64.3%
IONS ... 42 ... 33.3%
TRES ... 42 ... 0.0%
TIER ... 42 ... 50.0%
TREE ... 41 ... 78.0%
ERNS ... 41 ... 58.5%
SIES ... 41 ... 46.3%
SEIS ... 41 ... 56.1%
SATE ... 41 ... 70.7%
ETAS ... 41 ... 58.5%
SENA ... 41 ... 70.7%
ANES ... 41 ... 51.2%
RENS ... 41 ... 31.7%
RISE ... 40 ... 60.0%
SENE ... 40 ... 87.5%
ENES ... 40 ... 80.0%
REST ... 40 ... 55.0%
EATS ... 40 ... 52.5%
STEN ... 40 ... 57.5%
NETS ... 40 ... 67.5%
RETS ... 39 ... 41.0%
SEIR ... 39 ... 53.8%
ESSE ... 39 ... 84.6%
SEES ... 39 ... 66.7%
SNEE ... 39 ... 61.5%
ETEN ... 39 ... 53.8%
NETE ... 39 ... 59.0%
TENS ... 39 ... 69.2%
REOS ... 39 ... 0.0%
IRES ... 38 ... 50.0%
SENT ... 38 ... 76.3%
SANE ... 38 ... 71.1%
EASE ... 38 ... 60.5%
ERST ... 38 ... 0.0%
NOSE ... 38 ... 71.1%
REEN ... 38 ... 52.6%
SEAN ... 38 ... 76.3%
RAIS ... 37 ... 29.7%
SITS ... 37 ... 56.8%
TANE ... 37 ... 64.9%
ONES ... 37 ... 32.4%
TEES ... 37 ... 43.2%
SNOT ... 36 ... 63.9%
TONS ... 36 ... 66.7%
NIES ... 36 ... 41.7%
TOES ... 36 ... 38.9%
SEAS ... 36 ... 44.4%
SUET ... 36 ... 47.2%
NOTA ... 36 ... 44.4%
TONE ... 36 ... 83.3%
NITS ... 36 ... 50.0%
AITS ... 36 ... 36.1%
REIN ... 35 ... 54.3%
RITE ... 35 ... 68.6%
EANS ... 35 ... 42.9%
ONIE ... 34 ... 0.0%
ORES ... 34 ... 35.3%
TOSE ... 34 ... 38.2%
AINS ... 34 ... 64.7%
NATS ... 34 ... 38.2%
IOTA ... 34 ... 5.9%
TINE ... 34 ... 79.4%
NOES ... 34 ... 55.9%
ESTS ... 33 ... 51.5%
SNIT ... 33 ... 51.5%
NEAT ... 33 ... 39.4%
TINS ... 33 ... 54.5%
ANTS ... 33 ... 60.6%
TIRE ... 33 ... 84.8%
AINE ... 33 ... 36.4%
ARSE ... 32 ... 71.9%
SAIN ... 32 ... 56.2%
ONST ... 32 ... 0.0%
UTES ... 32 ... 0.0%
TEST ... 32 ... 50.0%
TETS ... 32 ... 75.0%
STET ... 32 ... 75.0%
EINA ... 32 ... 6.2%
SENSE ... 32 ... 75.0%
ESNES ... 32 ... 71.9%
USES ... 32 ... 43.8%
USER ... 32 ... 40.6%
TANS ... 32 ... 43.8%
EONS ... 32 ... 28.1%
TRIE ... 31 ... 64.5%
SANS ... 31 ... 35.5%
SUIT ... 31 ... 41.9%
ANTI ... 31 ... 48.4%
RAIT ... 31 ... 74.2%
TIAR ... 31 ... 71.0%
ISNA ... 31 ... 0.0%
RATO ... 31 ... 32.3%
SONE ... 31 ... 64.5%
RUES ... 31 ... 61.3%
ANIS ... 30 ... 53.3%
OUST ... 30 ... 73.3%
NOTE ... 30 ... 70.0%
SIST ... 30 ... 50.0%
SUES ... 30 ... 33.3%
TAOS ... 30 ... 23.3%
RASE ... 29 ... 62.1%
TORE ... 29 ... 69.0%
SEIL ... 29 ... 72.4%
OSES ... 29 ... 41.4%
LIES ... 29 ... 69.0%
SINS ... 29 ... 75.9%
CATE ... 29 ... 82.8%
ESSES ... 29 ... 96.6%
EARS ... 29 ... 62.1%
ONER ... 29 ... 37.9%
ERNE ... 29 ... 55.2%
REAN ... 29 ... 62.1%
RAIN ... 28 ... 60.7%
RANI ... 28 ... 57.1%
RENT ... 28 ... 71.4%
URES ... 28 ... 21.4%
SILE ... 28 ... 89.3%
SANT ... 28 ... 64.3%

[erakis17]

From about 450 games, here are the statistics for length of longest word in a board, found by my solver. The word list I'm using is SOWPODS (up to 15 letters) plus the long words from YAWL, which are a pretty good match based on what people like lalatan have posted in the favorite word thread. (Not perfect, though!)

LENGTH: 7   8    9  10  11  12  13  14  15  16  17  18  19  20  21  22  23
#TIMES: 20  82  68  33  16   7  15  100 58  38  10  4   6   2   0   0   1

The two-peaked structure (see attached graph) is pretty telling: when the board creation algorithm changed in 2015, the mix of boards was supposed to be 2/3 new (more long words), and 1/3 old. Here are the numbers of long words in the combined word list I'm using:

LENGTH:  14   15   16   17   18    19   20   21   22   23   24
#TIMES:9116 5757 3242 1796  968   508  244   99   40   22   15

The second peak is very consistent with the algorithm for new boards = "choose a random long word >= 14 letters, and then make an otherwise random board containing it".

[Spike1007]

That's pretty interesting. I didn't realize that there was a mix of new/old style boards here, but that really shows up in your plot. Also, your sample shows why there's so much clicking involved when looking for boards with existing long (20+ letter) words.

I've been a little puzzled (no pun intended) about how they build the boards. Your proposed algorithm (one long word plus random stuff) makes sense. I assume that the random part pays some attention to letter frequencies (kind of like real Boggle does), but it sometimes seems that I get multiple Qs, for example, more often than chance would dictate.

And then, of course, I assume that they filter the resulting boards to throw out those with certain prohibited words that might offend our sensibilities.