A general model of hippocampal and dorsal striatal learning and decision making

New version with corrections suggested by the ReScience paper's reviewers.

A general model of hippocampal and dorsal striatal learning and decision making

The source code associated with this repository allows to simulate the behavior of artificial rats in a variant of the Morris water-maze: the Pearce, Roberts and Good (1998) experiment. The agents' models consist in a reliability-based arbitration mechanism between an associative-learning strategy and a place-based navigation strategy. The associative-learning strategy is implemented using a classical Q-learning algorithm which is fed with visual input in an egocentric frame of reference. The place-based strategy on the other hand works in an allocentric fashion and is implemented using the Successor-Representation initially proposed by Peter Dayan (1993).

This code was created to conduct a partial reproduction and replication study of Geerts, Chersi, Stachenfeld and Burgess (2020), which addressed the Pearce, Roberts and Good (1998) in the beginning of their paper.

Geerts and colleagues provided a python implementation of their model and of the simulated task, accessible on ModelDB. Nevertheless, while inspecting the provided modules, we found several discrepancies with the original experimental protocol.

The implementation we propose in this article is a modified version of the original code of Geerts, Chersi, Stachenfeld and Burgess (2020). It
propose modifications of the code to more accurately simulate the original experimental protocol of Pearce et al. (1998), and perform new optimizations of model parameters in order to find those that enable the model to replicate rats behavior in this adapted version of the simulated task. The code was also modified to be far less computationally expensive (both in processing time and in memory consumption), while accepting new spatial navigation strategies and arbitration mechanisms, and finally to perform supplementary experiments which are unrelated to this paper. We find that the new version of the code enables the model of Geerts et al. (2020) to account for nearly all aspects of the rats original learning curves.