The role of reinforcement learning models to assess decision-making in the Iowa Gambling Task under the influence of alcohol

Abstract

Event Abstract Back to Event The role of reinforcement learning models to assess decision-making in the Iowa Gambling Task under the influence of alcohol Alvaro Guevara1* and Michael N. Smolka1 1 Technische Universitaet Dresden, Germany The Iowa Gambling Task (IGT) [1] is one of the most widely used laboratory tests to study decision making. In this task, participants attempt to maximize gains by selecting cards from four different decks, each of which is characterized by a specific frequency and magnitude schedule of rewards and punishments. Two of the decks are advantageous because they possess a positive average net payoff, whereas the other two have negative average net payoff, thus considered disadvantageous. A previous study [2] indicates that alcohol tends to improve performance in the task, that is, participants hit more often the advantageous than the disadvantageous decks after alcohol intake, but this effect was observed in combination with MDMA (amphetamines). Therefore, it remains an open issue to determine the specific decision making deficits elicited by alcohol consumption while completing the IGT. Moreover, the nature of the IGT (basically a four-armed bandit problem) makes reinforcement learning models suitable to assess such deficits. We conducted a study where 77 healthy participants performed the IGT, once sober and once alcoholized, following a randomized cross-over design. We fitted two reinforcement learning models to the data, namely, the Expectancy-Valence (EV) model and the Prospect-Decay-Independent (PDI) model (for more details on these models we refer the reader to [3]). We confirm the findings in [3] in the sense that PDI provides an overall better fit to the data compared to EV. Furthermore, we assess the role of alcohol intake in IGT performance by completing regression analyses where the alcohol concentration of the participant acts as an explanatory variable for the difference between the estimated values of the model parameters across sessions. Acknowledgements This research is supported by NIAAA Grant U01-AA017900, DFG Grant SM 80/7-1, DFG Grant ZI 1119/3-1, DFG Grant ZI 1119/4-1.