Andrea Zinno's Pages |
Starting with TooThought! 2.7, a new "Mistake Function" has been introduced. The goal is to make TooThought! able to play in a more human-like fashion, giving the possibility to make mistakes during its analysis process, choosing moves that are not necessarily the best.
The "mistakes" operates at board evaluation level: when TooThought!, during its analysis of the moves, has to evaluate a board (for example when reach the leaf of the tree representing the boards configurations that can be reached starting from the current one) then a mistake is introduced as a perturbation of the real board value. So the evaluation function, in general terms, become as follows:
value = F(board) + R(mistake)
where F represent the evaluation function (linear or neural) and R is a random value choosen according to the mistake strategy and parameters. At moment two different strategies are available:
In the first case the mistake is not related to the value of the board, in the second the mistake grow with such value. To be noted that in the first case a board can change its value from positive to negative ( and viceversa) only when the actual board value is less than X, while in the second when X is greater than 100%.
Testing the Mistake Function
The comparison is performed configuring TooThought! 2.7 to play with itself all the openings of length 6, with an analysis depth of 8 and solving the game at 18 empties.
The first table show the results, from a Black player point of view, when no mistakes are allowed:
The following tables, to the contrary, show the results related to different mistake strategies and values for the White player:
As You can see, as the mistake parametrs grow the White performance worsen.