We had to do something similar for our Intro to CS course at Tufts years ago: implement both a tic-tac-toe game and an algorithm that would play optimally. Second stage of the project was that, after a player's third move, the oldest "X" or "O" would disappear each turn. I have no idea if they still use that project, but the feeling when your computer player started working and you couldn't beat it was like nothing else, and was one of the reasons I went into CS rather than Economics.
I also had to implement a Tic Tac Toe "AI" when I was in college about a decade ago. Last year I was revisiting my college files backup and decided to upload it into GitHub:
One time I wrote something like this, and shared on reddit. Like you, I used minimax to solve tic tac toe. My comment was downvoted to oblivion and one highly upvoted person wrote:
> You don't need "AI" to solve tic tac toe, it's a solved problem!
Even though it truly is a solved problem, it's much more intuitive (and fun) to implement a tic-tac-toe solver from scratch using minimax than any lookup tables heuristic.
Since it is a generic solution, I believe the minimax solver can also be easily modified to implement an "AI" player for that Roman version of the game as well, where pieces are removed.
By the way, using minimax was also one of that college task requirements!
Brings back memories. I think implementing an AI for tic-tac-toe was one of my first complete implementations of a RL algorithm; circa 2008, I think :)
