Abstract
Choice behavior research accounts for temporal and spatial variables that mediate the relationship between environmental events and choice-making. As machine learning (ML) tools are increasingly utilized for behavioral data analysis, we evaluated the efficacy of various algorithms in retrodicting reinforcement contingencies from binary choice sequences. A single-neuron spiking neural network (S-SNN) model was recently used to retrodict which reinforcement contingencies were in effect during training, effectively performing a reverse inference from observed behavior to learning conditions. In this study, we assess the ability of various ML models to perform the same task: using nine 5-s windows of choice behavior following the delivery of one of nine reinforcers within seven components defined by concurrent variable-interval schedules sampled across 50 training sessions. We evaluated each model’s ability to infer the contingencies shaping choice behavior (i.e., learning histories). To enable direct comparisons with the prior S-SNN study, we used the same datasets. Our findings offer guidance for selecting ML tools suited to behavioral data and highlight the importance of modeling spatiotemporal structure when analyzing learning histories.