Wordle Solver

The most comprehensive, exhaustive, parameterized command-line wordle (https://www.powerlanguage.co.uk/wordle/) solver. Wordle is a really popular game made viral by it's inscrutable and quirky emoticon-based game description.

The solver boasts a 99.35%+ accuracy on the 8636 valid 5-letter words. Features:

Supports 4 modes: play, show (to show a solution for a specific word), solve (to solve a puzzle online) and eval (evaluate the performance)
Deterministic
Highest accuracy of all solutions evaluated
Fully tested
Latency ~0.26s per run Current dictionary used is the valid Scrabble dictionary.

Solver’s attempt to solve the Jan 10, 2022 wordle for the word query:

⬛🟨⬛⬛🟨 AROSE
⬛🟩⬛⬛⬛ TUMID
⬛⬛🟨⬛⬛ GLYPH
🟩🟩🟩🟩🟩 QUERY

Algorithm

With the settings for non-strict play, using positional

Find all candidates that fit the criteria
Amongst the valid candidates, compute a distribution of letters at each position
Find a word from all valid guesses which optimizes sum(P(letter at pos i)) + 0.5 * sum(P letter not at pos i)
Repeat

The 56 failure cases are sakes, mooed, jived, wanes, jocks, minks, wades, jaded, zoner, joker, wived, jakes, mozos, goxes, vills, rover, zooks, cozes, jibes, wakes, hajes, joked, sinhs, zaxes, yaffs, hiker, bases, moved, bises, zills, hided, eaved, vined, surfs, jiber, gibed, dozer, fuzed, mixed, boxed, waxes, waves, vomer, egged, mazed, pests, hived, socks, fazes, vests, jibed, mewed, hazes, sooks, woods, sinks For all these words, there are 2-5 candidate words left at the last guess, and with a random last guess, there is a probability of guessing these too.

Usage

Play it yourself

python main.py -m play

Guess? tares
TARES
⬛🟩🟨⬛⬛
Guess? unlit
UNLIT
⬛⬛⬛⬛⬛
Guess? raver
RAVER
🟨🟩🟩⬛🟩
Guess? favor
FAVOR
🟩🟩🟩🟩🟩
Solved! - favor

Solve for an unknown word

python main.py -m solve

Try the word [TARES]. There are 8636 possible words: ['aahed', 'aalii', 'aargh', 'abaca', 'abaci']...
How did it do (0=⬛, 1=🟨, 2=🟩) e.g. 00000? 00000
Try the word [NOILY]. There are 575 possible words: ['biddy', 'biffy', 'bifid', 'bigly', 'bijou']...
How did it do (0=⬛, 1=🟨, 2=🟩) e.g. 00000? 02002
Try the word [DHOBI]. There are 39 possible words: ['bobby', 'boggy', 'booby', 'boogy', 'boomy']...
How did it do (0=⬛, 1=🟨, 2=🟩) e.g. 00000? 20100
Try the word [DODGY]. There are 3 possible words: ['dodgy', 'doggy', 'dowdy']...
How did it do (0=⬛, 1=🟨, 2=🟩) e.g. 00000? 22122
Try the word [DOGGY]. There are 1 possible words: ['doggy']...
How did it do (0=⬛, 1=🟨, 2=🟩) e.g. 00000? 22222
Solved! = doggy

Show a solution for a specific word

python main.py -m show -w oozed

Word [OOZED]
Choosing [tares]. Total 8636 candidates: ['aahed', 'aalii', 'aargh', 'abaca', 'abaci']...
TARES
⬛⬛⬛🟩⬛
Choosing [coled]. Total 288 candidates: ['bedel', 'bedew', 'bevel', 'bezel', 'bided']...
COLED
⬛🟩⬛🟩🟩
Choosing [howdy]. Total 31 candidates: ['boded', 'boned', 'booed', 'bowed', 'boxed']...
HOWDY
⬛🟩⬛🟨⬛
Choosing [bipod]. Total 16 candidates: ['boded', 'boned', 'booed', 'boxed', 'domed']...
BIPOD
⬛⬛⬛🟨🟩
Choosing [dozen]. Total 8 candidates: ['domed', 'dozed', 'foxed', 'joked', 'mooed']...
DOZEN
🟨🟩🟩🟩⬛
Choosing [oozed]. Total 1 candidates: ['oozed']...
OOZED
🟩🟩🟩🟩🟩
Solved! - oozed
Woohoo! Solver solved it in 6 guesses!

Evaluate its performance

python main.py -m eval -k 1000

Evaluating on 1000 words
k=10:	Failed: 1	Accuracy:90.00%	Avg Time: 0.258s
k=20:	Failed: 1	Accuracy:95.00%	Avg Time: 0.250s
k=30:	Failed: 1	Accuracy:96.67%	Avg Time: 0.249s
k=40:	Failed: 1	Accuracy:97.50%	Avg Time: 0.244s
k=50:	Failed: 1	Accuracy:98.00%	Avg Time: 0.239s
...
k=970:	Failed: 10	Accuracy:98.97%	Avg Time: 0.236s
k=980:	Failed: 10	Accuracy:98.98%	Avg Time: 0.236s
k=990:	Failed: 10	Accuracy:98.99%	Avg Time: 0.236s
Failed on: ['jived', 'hides', 'razer', 'zooks', 'jills', 'gibed', 'wises', 'yipes', 'wipes', 'sises']
Distribution of remaining candidates: [(4, 5), (3, 3), (2, 1), (5, 1)]
K=999:	Failed: 10	Accuracy:99.00%

Run Tests

python -m unittest runs the entire test suite.

Evaluation

Using a dictionary of scrabble words, there are 172,819 total words and around 5% of them are exactly 5 letters long (8,636). The algorithm devised achieves a 99.28% success rate at guessing the right word, failing to get the correct the answer for 62 words.

Other settings achieved:

Global character frequency heuristic: Couldn't solve for 133 out of 1000 random samples (86.7% Success rate)
Conditional character frequency heuristic, on candidates left: Couldn't solve for 100 out of 1000 random samples (90.0% Success rate)
Non-strict solution: Couldn't solve for 46 out of 1000 random samples (95.4% Success rate)
Position-aware frequency heuristic + bug fixes: Couldn't solve for 9 out of 1000 random samples (99.1% Success rate)

Future Work

Support passing in solver settings through command line, including the weight of NON_POS_WEIGHT
Support passing in a custom dictionary through settings
Expose customizable number of letters in the word (N) and MAX_GUESSES.

Comments

Improve solver tests
This PR:

Renames one of the two tests named test_guess_next_word. Currently, the first test with this name is not getting run at all.

Adds a new test, test_guess_next_word_no_options, which ensures that we raise a no candidate exception if there's no suitable candidates (ie empty candidate lists or all candidates are already used)
opened by veeral-patel 3
Add CI + Makefile
This PR:

Adds CI to the repository using GitHub Actions. The CI pipeline runs unit tests and generates a code coverage report.

See a sample CI pipeline execution here:

https://github.com/veeral-patel/wordle-solver/runs/5270506138?check_suite_focus=true

Creates a Makefile which can be used to run tests and compute code coverage with make test and make cover respectively. This way no one needs to memorize any commands.

(Note you may need to install coverage with pip3 install coverage)

In addition, make cover-report opens a HTML UI with coverage data, broken down by line, as you can see here:
opened by veeral-patel 2
WIP: Add CI (using GitHub Actions)

Hi Deedy - was working on adding CI to the repository in order to run our tests (and help me get my feet wet with working with the codebase). Unfortunately looks like I created this pull request on your repo rather than my own fork. I'll get it working and let you know when I have something to merge :)

opened by veeral-patel 0
Wrong exception management

While trying to debug the code it didn't stop at the failing sentence. The problem is that main code catches all the exceptions and does nothing to either fix it or solve it.

https://github.com/deedy/wordle-solver/blob/deed27f02577aa3b00387287114b443f93ebbdc1/main.py#L72-L78

I think the code shouldn't catch exceptions if it's not going to do anything with them.

opened by Temporalin 0
Fails with non-English languages

I tried running the solver with a Spanish dictionary and it broke in this line: https://github.com/deedy/wordle-solver/blob/deed27f02577aa3b00387287114b443f93ebbdc1/game/solver/solver.py#L131

It happens when c='ñ', because ord('ñ') == 241.

opened by Temporalin 0