Do Human Reinforcement Learning Models Account for Key Experimental Choice Patterns in the Iowa Gambling Task?

Abstract

The Iowa gambling task (IGT) is widely used to study risky decision-making and learning from rewards and punishments. Although numerous cognitive models have been developed using reinforcement learning frameworks to investigate the processes underlying the IGT, no single model has consistently been identified as superior, largely due to the overlooked importance of model flexibility in capturing choice patterns. This study examines whether human reinforcement learning models adequately capture key experimental choice patterns observed in IGT data. Using simulation and parameter space partitioning (PSP) methods, we explored the parameter space of two recently introduced models-Outcome-Representation Learning and Value plus Sequential Exploration-alongside four traditional models. PSP, a global analysis method, investigates what patterns are relevant to the parameters' spaces of a model, thereby providing insights into model flexibility. The PSP study revealed varying potentials among candidate models to generate relevant choice patterns in IGT, suggesting that model selection may be dependent on the specific choice patterns present in a given dataset. We investigated central choice patterns and fitted all models by analyzing a comprehensive data pool (N = 1428) comprising 45 behavioral datasets from both healthy and clinical populations. Applying Akaike and Bayesian information criteria, we found that the Value plus Sequential Exploration model outperformed others due to its balanced potential to generate all experimentally observed choice patterns. These findings suggested that the search for a suitable IGT model may have reached its conclusion, emphasizing the importance of aligning a model's parameter space with experimentally observed choice patterns for achieving high accuracy in cognitive modeling.