Creating a Wordle Bot - Part 1 - Logic

I have been a fan of Wordle since it first appeared and I have been playing it on and off since then. NYT came up with a Wordle bot that tells you how good your solution was by considering how well your guesses would have performed against every possible word that satisfies the result from the previous step. 

I decided to try building this bot to improve my Python coding in conjunction with GPT 4o. To begin with, I found two lists of words

• All valid 5 letter words that can be used in the guesses

• All possible Wordle words

In terms of the algorithm, I decided to try the brute force method first. Here I would build a decision tree and come up with the optimal guess at each node based on the output from the previous step.

First, I would compare each guess word against the list of possible words and for each pair, create a 5 letter match string by comparing each letter of the guess word against the possible word and if the letter is present in the exact same spot, it is represented by “E”, if it is present but in a different spot, it is represented by “P” and if it is not present at all, it is represented by “N”. 

Using this string, I create a new dictionary per guess word which has the following elements:

• Unique match string

• Count of string occurrences

• List of possible words that generated that string

This is done by comparing the guess word to the possible word, generating the match string, then checking if that match string already exists in the dictionary. If so, then increment count by 1 and add the possible word to the list of words that generated that string. If it doesnt exist, create a new entry with that string, initialize the count to 1 and add the possible word.

So an entry might look like {"ENNNP", 3, ["crate", "chase", "crane"]} for the guess word "cello". Once I have gone down the list of every possible word for a given guess word, I would calculate average count of string occurrences for every guess word. Here my hypothesis is that the best guess is one which results in minimal number of possible words so no matter, what the word of the day is, it minimizes the number of words left at every stage.

However this alone may not be the best approach as the number of words left in subsequent nodes may be higher even if the initial average is lower. So I decided to take the 20 words with least average occurrence count, create the next level nodes for all 20, find average count across all the second level nodes and then choose the one with the lowest average. Then I drop the remaining 19 and stick to that one. 

As a validation for this approach, the initial list of 20 I found was: ['trace', 'crate', 'parse', 'slate', 'crane', 'heart', 'stale', 'least', 'react', 'cater', 'tried', 'clear', 'train', 'trade', 'alien', 'dealt', 'trice', 'chart', 'cleat', 'diner']. Since the Wordle bot always starts with ‘crane’, I am probably on the right track. Also these words contained the most possible output strings which makes sense as the best possible words would break up the universe of possible words into multiple paths thus reducing the number of steps needed in each path.

Once I have the first node, the list of branches going out are  represented by the dictionary I mentioned above. So if the word “crane” has 140 possible output strings, I would have 140 branches going out to the next node. Now in this next node, I repeat the exercise above where my guess word list remains the same but the possible word list has now shrunk to the list of possible words for that output string. I redo the whole exercise to generate possible output strings, their occurrences and the list of possible words.

Now one consideration here is that I only use the list of possible words for the incoming branch as the possible word list for that node which would be the equivalent of the “Hard mode” in Wordle but this would lead to some sub optimal solutions as words that don’t match the pattern could still be useful for eliminating possible words. So I am using the full list though that leads to additional computational overhead.

Now to implement this in Python, I needed to understand what data structures I can use to store the output at each step and possible how to represent the final tree. Also 4o suggested parallelization so that was something I incorporated. In the next part, we will jump into the code.