Foraging is an essential behavior for animal survival and requires both learning and decision-making skills. However, despite its relevance and ubiquity, there is still no effective mathematical framework to adequately estimate foraging performance that also takes interindividual variability into account. In this work, foraging performance is evaluated in the context of multi-armed bandit (MAB) problems by means of a biological model and a machine learning algorithm. Siamese fighting fish (Betta splendens) were used as a biological model and their ability to forage was assessed in a four-arm cross-maze over 21 trials. It was observed that fish performance varies according to their basal cortisol levels, i.e., a reduced average reward is associated with low and high levels of basal cortisol, while the optimal level maximizes foraging performance. In addition, we suggest the adoption of the epsilon-greedy algorithm to deal with the exploration-exploitation tradeoff and simulate foraging decisions. The algorithm provided results closely related to the biological model and allowed the normalized basal cortisol levels to be correlated with a corresponding tuning parameter. The obtained results indicate that machine learning, by helping to shed light on the intrinsic relationships between physiological parameters and animal behavior, can be a powerful tool for studying animal cognition and behavioral sciences.