Georgia Institute of Technology
Comparing Computational Models of Rat Decision-Making in a Three-Armed Restless Bandit Task
Abstract
dc:description.abstractWe compared four computational models of rat decision-making in a three-armed restless bandit task: Reinforcement Learning with a choice kernel (RL), Active Inference (AIM), Win-Stay-Lose-Shift (WSLS), and a Foraging model. Models were evaluated using yoked simulations, where the model receives the rat's actual choices and rewards up to each trial and predicts the next choice. RL fit the data best (mean Hamming similarity 0.60, winning 59% of 110 sessions), though AIM was a close second (0.58, 28%). RL's fitted parameters showed low Q-value learning rates and high choice kernel weights, suggesting the model predicts choices primarily from recent choice history rather than learned values. RL parameters clustered by rat identity while AIM parameters clustered by walk schedule, indicating that RL captures individual behavioral differences and AIM tracks task difficulty. AIM's exploration weight collapsed to around zero in most sessions. All four models failed to capture sustained two-arm alternation (rat streaks up to 87 trials vs. 7 for RL) and systematic arm preferences that differed across rats on the same walk. RL with a choice kernel provides the best available account of rat behavior in this task, but there are spatial and motor components of decision-making that are still unexplained by the current models.
Degree
thesis:*- Name thesis:degree_name
- Computational Science and Engineering, MS
- Department thesis:degree_department
- Computational Science and Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Drury, MacKenzie
- Advisors dc:contributor.advisor
-
- Sederberg, Audrey
- Wang, Kai
- Committee member dc:contributor.committeemember
-
- Wu, Anqi
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81597