Skip to Main content Skip to Navigation
Theses

Computational learning noise in human decision-making

Abstract : In uncertain and changing environments, making sequential decisions requires analyzing and weighting the past and present information. To model human behavior in such environments, computational approaches to learning have been developed based on reinforcement learning or Bayesian inference. To further account for behavioral variability, these computational approaches assume action selection noise, usually modeled with a softmax function. In the first part of my work, I argue that action selection noise is insufficient to explain behavioral variability and show the presence of learning noise reflecting computational imprecisions. To this end, I introduced computational noise in the standard reinforcement learning algorithm through random deviations in the noise-free update rule. Adding this noise led to a better account of human behavioral performances in reward-guided tasks (Findling C., Skvortsova V., et al., 2018a, in prep). The presence of learning noise led me to investigate whether this noise could have a functional role. In the second part of my work, I argue that this learning noise actually has virtuous adaptive properties in learning processes elicited in changing (volatile) environments. Using the Bayesian modeling framework, I demonstrate that a simple learning model assuming stable external contingencies with learning noise performs virtually as well as the optimal Bayesian adaptive process based on inferring the volatility of the environment. Furthermore, I establish that this learning noise model better explains human behavioral performances in changing environments (Findling C. at al., 2018b, in prep).
Complete list of metadatas

Cited literature [416 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-02612235
Contributor : Abes Star :  Contact
Submitted on : Tuesday, May 19, 2020 - 6:46:02 AM
Last modification on : Wednesday, October 14, 2020 - 4:06:39 AM

File

FINDLING_Charles_these_2018.pd...
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-02612235, version 1

Citation

Charles Findling. Computational learning noise in human decision-making. Cognitive Sciences. Sorbonne Université, 2018. English. ⟨NNT : 2018SORUS490⟩. ⟨tel-02612235⟩

Share

Metrics

Record views

83

Files downloads

26