What to do when a Monte Carlo tree search takes up a memory limit

Question

I recently showed interest in finding monte carlo games.

I have read several works, but I use the "Search the Monte Carlo Tree". Doctoral dissertation on Chaslot, G, as it’s easier for me to understand the basics of finding a monte carlo tree

I tried to encode it and was stuck in a specific problem. The algorithm tries to deploy one node in the game tree for each simulation. This quickly escalates into a memory problem. I quickly read the article, but it doesn't seem to explain what this technique will do if it reaches a certain memory limit.

Can you suggest what this technique should do if it has exceeded a certain memory limit?

you can look here: http://www.unimaas.nl/games/files/phd/Chaslot_thesis.pdf

Answer 1

One very effective approach is to grow the tree more slowly. That is, instead of expanding the tree every time you reach the leaf node, you expand it when it has at least k visits. This will significantly slow down the growth of the tree and often does not reduce productivity. One of the authors of the Fuego Go program told me that he tried the approach, and in practice it worked well.

This idea was originally described in this article:

Remy Cool Efficient selectivity and backup operators in the search for the Monte Carlo tree. In the Computers and Games Section, pp. 72-83. Springer, 2007.

It has also been used in:

Max Roshke and Nathan Stertevant. UCT improvements in Chinese shahs using the Endgame database, IJCAI Workshop on computer games, 2013.

Answer 2

You can discard all nodes with the number of visits that are less than any threshold that has not been visited recently (how many times have been played back). This is a quick but not effective solution. It is also better to implement gradual expansion.