The Wordle Bot Cheats
May. 20th, 2022 07:59 pm Well, not precisely. But it's method of play would count as cheating for a human, and its measure of skill is based on a "cheating" algorithm.
The Wordle Bot internalises the finite set of all the answers set for the game. Note - and this is important - that this is not a complete set of the five-letter words in English, nor even a complete set of the common five-letter words: it does not, for example, include slats or thine.
At any stage it analyses the finite set for the pattern which will eliminate the most possibilities by what it will include/exclude. (If aeiou were a word, guessing it would allow you to cross off all the words which use letters which it does not use and all the words which do not have green letters in the same place.) Given the size of the set, it can brute-force this; I suspect that it actually internalises the number of words left after each valid word as a starting guess at the beginning of the day, so that it can provide that feedback to the user quickly.
The fact that it always recommends "crane" as an opening word reflects a brute-force analysis. A guess merely based on letter frequencies would omit c and probably include t. Later and rants would be better guesses based on general letter frequencies.
After each guess, it partitions the set into the remaining finite subset of the starting finite subset for that guess which match all the known conditions. It calculates the number of gusses left in the list if each word were chosen, and selects the choice with the smallest value.
If you have guessed a valid English word which is not in the complete set of Wordle answers it assigns that guess a skill of zero. What it really means is that the user has a bigger vocabulary (which should be a plus, not a minus, in analyzing skill).
There are two points to make:
1) This is effectively cheating. By doing a brute-force analysis against the internal list kept by the game, it performs the equivalent of looking in the back of the textbook for the answer.
2) This is also not paying a game. As with any other game, just using brute force to calculate one's moves - what it effectively recommends as strategy - is not in any meaningful sense (other than the von Neumann and Morgenstern one) a game.
Expected play behaviour is for players to bring their general prior knowledge, in this case a general knowledge of the English language and not of the arbitrary subdomain which is "the set of all Wordle answers".
To approach it in a fair manner it should use as a basis all the five-letter words in the OED. Heuristics allowing solutions would have to take into use the probability that a word is current enough to be considered. (Thine I use every week; I haven't seen hight ("named", not misspelled height) in the wild in current use ever; lossy is a technical term with limited use, etc.)