Question about the Monte Carlo AI

by **jeffwadsworth** » 02 Mar 2012, 21:26

Cards in the past have been denied because of AI issues. Case in point, any card that allows one to search their library. Doesn't the Monte Carlo AI help to get around issue like this one...at least in regards to massive trees, etc.

by **Tyrael** » 03 Mar 2012, 12:09

I am not a pro regarding AI issues, however, as far as I understood, the MCTS AI randomly plays virtual games and rates the outcome of each game to get the best choice at the current game state. Thus, in order to have it make a good decision in selecting a card from its library, it would have to play random games for each card from the library to decide, whether that card is the best option or not, this would probably take far too much time, as far too many possible trees exist in that case.

So as far as I can tell, the answer to your question is no, the MCTS does not get around this issue.

by **Void** » 01 Aug 2012, 19:25

Does MCTS require the whole game state to be copied between nodes, or just the information about the move is sufficient?

by **pcastellazzi** » 01 Aug 2012, 23:03

jeffwadsworth wrote:Cards in the past have been denied because of AI issues. Case in point, any card that allows one to search their library. Doesn't the Monte Carlo AI help to get around issue like this one...at least in regards to massive trees, etc.

Tutoring (cards that search cards) is a work in progress. I did some of the work required to make it functional with Melvin's help. I will continue to work on it as soon as i have some time to spare.

by **melvin** » 02 Aug 2012, 05:18

Void wrote:Does MCTS require the whole game state to be copied between nodes, or just the information about the move is sufficient?

Each node stores the move that lead to the node.

The way we've implemented it is that each simulation starts with a copy of the current game state, which is modified as it moves down the tree.