Next bottlenecks and improvements:

1) My agent is not working when I submit it to Kaggle, taking too much time
2) I am thinking of using RL now to solve this problem better

Solution: To calculate scores of all the daughter nodes before updating the root node at once, make alpha beta for each possible move independent of each other by removing their global scope, use hard-coded scores for each move, and simulate for the root node

If you are not updating the alpha beta scores of the root node, then the alpha beta scores of one branch shouldn't flow to its sister branch, because the root node that used to connect them isn't in the update pipeline

Error: Each branch evaluation updates the global alpha beta variables, which I then use for a new branch evaluation, but without updating the root node alpha beta scores

And this is silly!!!