Modeling social conventions with Sequential Episodic Control

López Hernández, Adriana

Modeling social conventions with Sequential Episodic Control

License

Citation

López Hernández, A. Modeling social conventions with Sequential Episodic Control. 2022. handle: http://hdl.handle.net/10230/54297

This citation was generated automatically.

Abstract

Computational models can help bring light to the underlying cognitive mechanisms responsible for the emergence of conventions in human societies. In order to pro- vide meaningful theoretical insight, these models of behavior should aim to resemble human-like performance. Deep Reinforcement Learning (deep RL) algorithms fail to do so; these techniques are data inefficient and require several training instances to approximate the speed of human learning. Episodic Reinforcement Learning (ERL) algorithms, for their part, seek to improve deep RL algorithms by implementing memory buffers that can counter the sample inefficiency problem. Nonetheless, this approach also falls short because it considers memories as isolated, discrete events. On the other hand, episodic control models provide a model-free, non-parametric ap- proach capable of rapid learning, thus resembling more closely human performance. These algorithms introduce a fast memory system inspired by the hippocampus that allows them to search for a solution without having to make strong assump- tions about the world. The present work will test the adequacy of one particular episodic control algorithm, the Sequential Episodic Control (SEC) model to sim- ulate human behavioral data in a repeated coordination game. This project will compare the modeled and behavioral data concerning efficiency, fairness, and sta- bility measures to evaluate the model’s performance. Finally, given the structure of the model, this project will examine the potential theoretical implications of human conventionalization, as well as the limitations and future work on this approach.