Symposia
Transdiagnostic
Peter F. Hitchcock, Ph.D.
Brown University
Providence, Rhode Island
Michael Frank, Ph.D
Professor
Brown University
Providence, Rhode Island
We are often told to think harder, pay attention, and not sweat the small stuff — but how do we ingrain such mental (cognitive) actions through reinforcement learning? This question is relevant to cognitive-behavioral therapies (CBT) that adopt a functional analytic approach to mental behavior (i.e., they seek to replace maladaptive mental actions with adaptive ones that fulfill the same function). This includes therapies targeting repetitive negative thinking (RNT), such as, rumination-focused CBT (which seeks to replace rumination with, e.g., active problem solving) and mindfulness-based treatments for generalized anxiety disorder (which seek to replace worry with, e.g., mindfully attending to the breath). Critically, most of what we currently know about reinforcement learning comes from the study of external (overt) actions; our lack of knowledge about how mental actions are learned fundamentally limits the precision with which therapies can target these actions. To overcome this challenge, we developed a novel paradigm, The Cognitive Actions Task. The task has two conditions: in the Cognitive condition, responses require taking the sum or difference of two numbers; in the Overt condition, responses simply require responding with key pairs at the top or bottom of the screen. The contingencies are otherwise matched; hence, the conditions differ only in that the former uniquely requires performing a mental operation (arithmetic). Consistent with predictions, in an adult sample (n=60), accuracy was lower in the cognitive condition in a learning (β (SE) = .25 (.07), p < 1e-3) and test phase (β (SE) = .50 (.22), p = .03). Performance also decreased as a function of (z-scored) delay between choices, with no significant interaction between delay and condition (condition β (SE) = .18 (.05), p < 2e-16, delay β (SE) = -.14 (.05), p < 1e-3; delay*condition p</em> > .38); this suggests that the working memory demands of holding an action in mind over time impose an additional (largely orthogonal) cost to that imposed by condition. In a subsequent Generalization test phase, we found that choices could be predicted by subtle differences in reward history, suggesting that a reinforcement-learning model can appropriately model this task. At ABCT, we will present these behavioral results alongside a mechanistically specific reinforcement-learning computational model. We believe that our approach can offer insight into the unique challenges of cognitive-action learning — with wide-ranging implications, including to cognitive-behavioral therapies for RNT.