To ensure that researchers gain the greatest recognition from the accelerometers, they tested a suite of machine learning algorithms (logistic regression, support-vector machines, random forests and gradient boosting machines). For each algorithm, they used a training and testing set which were built from captive experiments, where each behaviour was manually coded from video of the animal wearing the accelerometer. Using the training set each algorithm was taught the pattern of each behavior. Then the algorithm was used to predict behaviours from the test set. This resulted in accuracies of over 80%.