Active Player Modeling in the Iterated Prisoner’s Dilemma

Table 3

Differences in average accuracy of active- and random-sampling approaches.

Player type	Difference in training accuracy	Difference in testing accuracy
AllC	0.001	0.001
TFT	0.061	0.150
NoisyTFT	0.018	0.123
Major	−0.014	−0.081
Major 5	0.019	0.101
Pavlovian	0.042	0.205

Although the differences in the accuracy of the active- and the random-sampling approaches were minimal in the training, large differences were observed in the testing. This implies that active-sampling may prevent overfitting (difference = accuracy of active-sampling − accuracy of random-sampling).