In terms of cognitive development, Piaget talks about concrete operations and formal operations. Here, playing out variations on the board (or screen) involves concrete operations and playing out variations in your head involves formal operations. Just as concrete operations normally precede formal operations in cognitive development, in learning go you can use concrete operations as a stepping stone to formal operations.
But, IMO, using undo against a bot to do that is not so good, because the bot chooses one and only one response. First, you want your play to be good against any response, and to learn about the position you face you have to explore a number of replies. Second, the bot chooses the response (or responses if it incorporates randomness) instead of you. You do not get the benefit of coming up with a set of possible responses yourself. Unless you are able to do exhaustive brute force search, you have to select which moves to try. That is an important go skill in itself. If you let the bot choose the responses, you lose opportunities to develop that skill.
So here is what I suggest. Do not use undos against bots. Instead, use another program, like an SGF editor, to explore variations until you feel you understand a position well enough to make a play, and then make the play and stick to it. It is important to stick to it, so that you get a sense of when you have explored a position well enough. As you use this method, you will find that more and more you will see variations in your head and will rely less and less upon playing them out in the other program. You will develop your ability to select possible plays and replies and will gain confidence in that ability, as well as in your ability to play out variations in your head and in your ability to judge the results.
Good luck!
