Abstract
<p>Attentional decision models (ADMs), including attention-modulated decision models and joint models of attention and decision making, are important tools for studying the cognitive processes underlying decision formation. However, these models often cannot be fit or compared using likelihood-based methods. Building on simulation-based and machine-learning approaches, this paper develops and tests an end-to-end neural network framework for parameter estimation and model comparison, tailored to the high-dimensional, multimodal structure of attentional decision data. In attentional drift diffusion models where likelihood-based fitting is feasible, we compared machine-learning methods to likelihood-based methods on parameter recovery, model recovery, and empirical data fitting. The machine-learning approach performed as well or better than the likelihood-based approach. The two methods yielded highly consistent conclusions for real participants on these tasks. In more complex models where likelihood-based methods are inapplicable, the machine-learning approach still achieved high accuracy in parameter recovery and model recovery. Moreover, we compared different feature-extraction strategies and network architectures and found that even the simplest were sufficient for reliable estimation and comparison with large-scale simulation data. Overall, our approach provides practical methodological tools that facilitate the advancement of attentional decision models and the development of associated theories.</p>