A python-based solver for NYT's popular Wordle game
I wrote this, so it's not cheating if I use it, but it is if you do! Also credit me if you use this code for anything other than personal puzzle-solving use.
This repository can be cloned by git clone... or you can just copy and paste it.
You need python 3 installed, that's non-negotiable.
Install the following packages to build the dictionary:
pip install nltk
python -m nltk.downloader corpus
python -m nltk.downloader brown
pip install math
Invoke the python code from your terminal of choice assuming you have python installed:
python wordle_helper.py
I don't currently care enough to make it an installer, etc, and I think this code's as good as it's getting
Python's nltk dictionary is not perfectly symmetrical with the Wordle author's dictionary, so not all games are solvable. It doesn't know the words gamer or fewer, from games in the last month or so as I've been testing it on my own.
First, the code uses the dictionary you installed in the download step to get a list of 8497 distinct five-letter words
For the initial move I suggest raise, which I've determined is the best first move. It was a very computational process for the same thing every time, so it's hardcoded.
A word is suggested by the algorithm
The user then enters this word into the game
The game then returns a five-color sequence (where each letter becomes green or yellow or remains black)
Using the code's shorthand the user enters this color data
Based on the color data the algorithm can narrow down the list of valid words
From there a new best word is determined from the remaining list of words
If it sounds subjective, think again.
This code is a much more advanced, resource-intensive version of a binary search algorithm with words.
The code goes through the list of possible words twice, once for guesses and once for answers. In each case, the answer gives a color response that can narrow down the list. We can't know what the answer is for a long list, but we can pick the word that will, on average, narrow down the list the most.
Luckily since the data we give and receive is much more complicated than a simple binary search, we can narrow down the list of possible answers much faster than just in half. Often best plays cut down lists by 90-95%.
Unfortunately this solution is O(n^2). For large cases it becomes slow. My workaround was to take a bunch of random subsets and play those and look for the most common winner among them. Still slow, but in a much more reasonable time-scale for the impatient human player.
In rare cases the remaining lists of words will not be effectively narrowed down by picking the best among the list, consider the following difficult remaining word list:
bould, could, gould, hould, mould, nould, tould, vould, would
In this case picking any one word would only eliminate itself. The algorithm can detect this! We can determine the bit value of the best play by how comparing the logarithms of input and mean output lists. My code's rule of thumb is that the guess should eliminate at least 1.2 bits of data, which means reducing the list by more than 57% of its previous length. That value was slightly hand-tuned.
Anyway, if we find that the best play from the list has a terrible expected bit delivery, we instead look back at the whole corpus of words and see if it can do better. In the above case it'd suggest something like chant which could act as a determinant on could, hould, nould and tould, either determining if any of them are the answer or eliminating them and leaving only five words left over from the original nine. In the latter case a different determining word would have to be found, if it's not the last round.
$ python wordle_helper.py
building dictionary...
Guess # 1
For your first guess, try 'raise'
Input a 5-letter sequence following based on your results:
g = letter is green
y = letter is yellow
b = letter is black
xxxxx = word is invalid
yyybb
narrowning down word list
out of 8497 words, only 89 remain
calculating next best play...
Guess # 2
next guess: glair
Input a 5-letter sequence following based on your results:
g = letter is green
y = letter is yellow
b = letter is black
xxxxx = word is invalid
bbgyy
narrowning down word list
out of 89 words, only 10 remain
remaining words: ['diary', 'tiara', 'ziara', 'acari', 'ajari', 'inarm', 'urari',
'fiard', 'arati', 'izard']
calculating next best play...
Guess # 3
next guess: ziara
Input a 5-letter sequence following based on your results:
g = letter is green
y = letter is yellow
b = letter is black
xxxxx = word is invalid
xxxxx
I'll remove that one and find the next best
narrowning down word list
out of 9 words, only 9 remain
remaining words: ['diary', 'tiara', 'acari', 'ajari', 'inarm', 'urari', 'fiard',
'arati', 'izard']
calculating next best play...
Guess # 3
next guess: tiara
Input a 5-letter sequence following based on your results:
g = letter is green
y = letter is yellow
b = letter is black
xxxxx = word is invalid
ggggg