Computational modeling of reinforcement learning using probabilistic selection task and instructional probabilistic selection task

Abstract

Introduction

Humans learn how to behave both through rules and instructions as well as through environmental experiences. It has been shown that instructions can powerfully control people's choices, often leading to a confirmation bias.

Aim

To compare learning parameters in reinforcement learning task with and without instructions.

Methods

We recruited 52 healthy adult control subjects (21 males, 31 females, age 30 ± 6.5 years). Participants completed Repeatable Battery of Neuropsychological Status (RBANSS). Twenty-seven participants completed additionally Probabilistic Selection Task (PST) while twenty-five participants completed Instructional Probabilistic Selection Task (IPST). To analyze learning parameters, we used Q-learning model with 3 parameters: learning rate due to positive and negative reinforcements as well as exploration-exploitation parameter.

Results

Both groups did not differ with respect to cognitive functioning measured with RBANSS (immediate and delayed memory, visuospatial abilities, language and attention); however, participants who completed PST had trend-level statistically faster learning rates due to positive (P = 0.099) and negative reinforcements (0.057) in comparison to participants who completed IPST. Both groups did not differ with respect to exploration-exploitation parameter (0.409).

Conclusion

In healthy adults, interference of confirmation bias can influence learning speed independent of cognitive functioning (immediate and delayed memory, visuospatial abilities, language and attention).

Disclosure of interest

The authors have not supplied their declaration of competing interest.