This is an archived copy of the original site of Marc Mandt

Decision Trees


The goal of a decision tree is to investigate the possible game outcomes given the current state of the game. In order to do this, a tree of game states is built where the depth of the tree represents how many moves into the future is currently being investigated. In order to evaluate the strength of one move branch over another, we return to the general strategy of gaining irreversible board locations.

Somewhere in the decision tree's branches, there exists the best possible outcome as well as the worse outcome and everything in between given any particular move. Since we are interested in gaining positions that we can't lose, it's the worse possible outcome as the result of each of immediately possible moves that will receive our attention. More to the point, if we choose the move that has the best worse possible outcome each time, i.e. creating irreversible pieces, then we are constantly making moves that improve our strategic strength.

In order to evaluate the value of each possible outcome [node] in the tree, one should use a board location value matrix. The behavior of the decision tree is further increased if a relative strength (player 1 vs. player 2)  is used as an indicator instead of the board value for the active player alone. Hence, a location value ratio is a great way to evaluate a tree node's weight:

Location Value Ratio: The sum of the active player's board values divided by the sum of the opponent's board values, where the each players values are established using their relative board location value matrix.

Parsing

It is necessary to parse these decision trees using some hueristic(s) so that they can be used real time.


< previous


Copyright© Marc Mandt, 2001
mmandt@mindspring.com