"Are you going to look?": Predicting anticipatory saccades towards future action outcomes via a neuroevolutionary machine learning algorithm

Kurzfassung/Abstract

When their actions are going to cause predictable effects in their surroundings, participants already move their eyes towards the locations of these future effects in anticipation. In this study, we used the neuroevolution of augmenting topologies (NEAT) algorithm to predict whether participants perform an anticipatory saccade (AS) towards the location at which a visual effect of their action is going to appear in the future. Previous research has linked such ASs to proactive effect monitoring (PEM) processes which facilitate the later comparison of expected and actual effect. Considerable inter-individual differences in the frequency and distribution of ASs, however, raise the question whether PEM also underlies the individual distribution of ASs across trials. We used data of two experiments and selected the data of participants with high proportions of ASs (i.e., presumed high PEM; criterion-based model) to train a population of neural networks (NNs) via the NEAT process to predict whether an AS was performed in each trial. We then assessed the performance of these NNs on the remaining data (testing) and on data from another experiment (transfer). Prediction accuracy of our criterion-based model was comparable to NNs trained on a randomly selected, bigger subset of data (basic model). Moreover, prediction accuracy for individual participants correlated with their scores on the criterion. These findings suggest that a systematic PEM process underlies the distribution of AS even though there is substantial inter-individual variation. More generally, we demonstrate the potential of using neuroevolutionary machine learning in theory-driven research.