Analytics Simulator for Wordle Game

For fun, wondering how I might use the popular Wordle game to learn some machine learning. Maybe implement a reinforced learning RL algorithm to learn to play? Monte Carlo method? Brute force?

Also some general game plat strategy questions:

What is the best start word?
what about a second start word if no letters are in the first word?

Some Analytics

The first brute force method was to try each word form the dictionary as the start word. Then play 2000 games using that start word, and save the win rate. After all that we got these top twenty words:

Start Word	Win of 2000	Win %
clump	1734	87%
clamp	1721	86%
plant	1719	86%
blimp	1718	86%
clint	1715	86%
chomp	1714	86%
mulct	1713	86%
paint	1711	86%
blurt	1711	86%
plumb	1710	86%
grump	1708	85%
nymph	1706	85%
wroth	1705	85%
bract	1703	85%
amble	1703	85%
thump	1702	85%
thank	1702	85%
trump	1701	85%
ponit	1701	85%
pinto	1701	85%

This is just from one round, and due to the nature of randomness this this will change at each run. But it's an interesting start.

Wordle Game Env

The game is simulated in /wordle_env. It starts a game, picks a random word or uses on given to it. Then returns a 5 x 6 matrix after each word attempt.

[
  [1, 1, -1, -1, 1],
  [1, 1, 1, 5, 1], 
  [5, 5, 5, 5, 5], 
  [0, 0, 0, 0, 0], 
  [0, 0, 0, 0, 0], 
  [0, 0, 0, 0, 0]
]

0 for not attempted yet
-1 for the letter that is not in the word
1 for the letter that is in the word but not that location
5 for the letter that is in the right spot

Brute Force Solver

First I created a simple brute force solver (in /solver/brute.py) that tried a word, then eliminated the words from the dictionary based on the environment state returned. From those words left it picked a random one and repeated.

Dictionary

For this I started with /usr/share/dict/words and then tired http://www.gwicks.net/dictionaries.htm, specifically usa2.txt but any dictionary will do.

What's Next