The highest scoring state from the outputs from the neural network was taken, and the score of the next highest predicted state subtracted:
This is similar to the method exploited by PHD [17].