MCTS AI development for "Tintas"

Monte Carlo Tree Search
ddyer
Posts: 23
Joined: Tue Feb 12, 2008 7:35 pm
Location: Los Angeles

MCTS AI development for "Tintas"

Postby ddyer » Wed Oct 18, 2017 12:23 am

Tintas is an abstract strategy game involving collecting colored tokens.
https://boardspace.net/tintas/english/tintas-rules.html

If you were trying to design a game to be a perfect vehicle for MCTS, this would be it. There
are a maximum of 49 moves in the game, and usually the branching factor is 6 or less. The
initial MCTS version, allowed 10 seconds per move, proved to be essentially unbeatable. The
AI problem was therefore to make a robot weak enough to be fun to play.

The usual first recourse is to reduce thinking time, but for Tintas even a minimal 1 second/move
thinking time was still very strong. The next recourse is to randomize the winning move based
on the win rate. This had the desired effect, but tended to make the robot look careless near the end;
Once all the variations are clearly losing, picking a random one looks like disinterest.

The final tweak, slightly evil, is to randomize only in winning positions. In losing positions the
robot makes the best move available, which means it fights to the end, and will more readily take
advantage of human mistakes that give it a chance to recover.
my game site is http://Boardspace.net

Return to “MCTS”

Who is online

Users browsing this forum: No registered users and 1 guest