2
$\begingroup$

AIs that rely on MCTS - like AlphaGo - create their decision tree as the game progresses. Do they start from scratch each game and build a new tree or do they keep the tree and grow it from game to game further?

Besides the possible limitations of storage space for the search tree, I don't see any obvious drawback in keeping and growing the tree, which seems to me to be the preferred option. Are there other reasons to start from scratch each game?

$\endgroup$
  • 3
    $\begingroup$ Is this just another random downvote again or is there something wrong with my question? As always, I'm happy to improve it, if you tell me what's the issue. $\endgroup$ – Demento Oct 22 '17 at 12:14
2
$\begingroup$

Your intuition is right. The tree is grown after each game.

Before the first play through, the values of each decision point are initialized (randomly or to some constant). Then at the end of one play through, the weights of the decision points are updated using Monte Carlo methods. In this way, on the next play through, the updated weights help the agent make a decision in the next game.

As for how the decision points are added to the tree, it depends on the application. For simple games like TicTacToe, we know all of the decision points. For big games, there are methods to trim the tree of bad performing branches.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.