Technical Program

Paper Detail

Paper: PS-2A.9
Session: Poster Session 2A
Location: H Lichthof
Session Time: Sunday, September 15, 17:15 - 20:15
Presentation Time:Sunday, September 15, 17:15 - 20:15
Presentation: Poster
Publication: 2019 Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany
Paper Title: An Active Inference Perspective on Habit Learning
Manuscript:  Click here to view manuscript
License: Creative Commons License
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Authors: Sarah Schwöbel, Dimitrije Markovic, Stefan Kiebel, TU Dresden, Germany
Abstract: When pursuing goals, agents choose actions according to a balance of two opposing systems: The goal-directed system which is slow but flexible, and the habitual system, which is fast but inflexible. It has recently been argued, that this dichotomy maps onto value-free and value-based decision-making processes. Here, we propose a hierarchical Bayesian cognitive model resting on active inference where habits correspond to adaptive prior beliefs over policies (action sequences). The policy prior is learned over time, dependent on the history of past actions, and enables the agent to dynamically arbitrate between the two systems when choosing actions. We show here that when an agent forms habits in a stable environment, habit formation leads to an increased performance and reduces the decision noise. In contrast, in a dynamic environment, habits might lead to maladaptive behaviour for specific free model parameters. This interaction between environmental properties and the agents generative model explains when and how habit formation is useful and when it can lead to aberrant behaviour.