Graphics Reference
In-Depth Information
The “winning potential” of the cell, the number of CPU tokens currently on the
board that share winning combinations with the cell, is weighted so the algorithm
favors cells that contribute to a win.
The “blocking potential” of the cell, the number of player tokens currently on
the board that share winning combinations with cell
n
, is unweighted so that the
algorithm pays less attention to a block, but still counts it in the overall priority
number.
8.3 Experimental Evaluation
In this section, we present the results of experimental determination of the optimal
cell detection box size in the vision block and evaluation of the two AI approaches
described above.
8.3.1 Vision Experiments
Although we tried to minimize detection error in our vision block, the performance
was still less than perfect. We found the major culprits to be the following:
If outer cell boxes are too small, they sometimes cut off parts of the cell content
If outer cell boxes are too big, they can accidentally capture parts of the grid,
causing incorrect results
In order to determine the optimal outer cell size, we devised an experiment that
would find what cell detection box size would yield the most accurate results. For
each of five different scale factors (the center cell scaled up by 1, 1.125, 1.25, 1.375,
or 1.5) we made a separate app and tuned its thresholds on a set of five tic-tac-toe
boards written in different styles and handwritings. Then we used additional ten
tic-tac-toe boards, not seen by the app during the tuning process, to evaluate the
detection error as a function of the scale factors.
In Table
8.1
, we present the number of errors for each cell size: V1, V2, V3, V4,
and V5 denote the app version with outer cell scale factors 1, 1.125, 1.25, 1.375, and
1.5, respectively.
We see that the scaling factor of 1.25 used in V3 can be chosen as the optimal
value which is neither too big nor too small.
Table 8.1
Vision block
accuracy with different cell
sizes
Number of errors
V1
V2
V3
V4
V5
X detect
2
0
0
4
2
O detect
0
1
0
4
2
E detect
0
0
0
0
0
