Cool, but assumes all words have an even probability.

Vs a week player the ideal word list should only include 'common' words.

Vs an ideal player you need to assume they will pick from the list of words that you least likely to guess using optimum play.

Generically, optimum play ends up with a somewhat random list, aka 90% of the time pick E, 10% of the time pick I etc.

PS: Actually generating this list is a 'hard' problem and vary dictionary dependent, but you can probably get reasonably close using some sort of genetic algorithm and a enough simulation time.