Convergence to a Common Protocol in Emergent Communication

Table of Contents

• About
• Getting Started
• Usage
• Directory Structure

About

Emergent Communication is a flexible, bottom-up framework that studies the properties of protocols created by artificial agents (sender-receiver pairs) to coordinate and solve a task. This project focuses on studying how individual idiolects formed through local interactions are extrapolated to form a communal language.

This codebase implements EC from scratch in a population, using the Gumbel-Softmax Relaxation. You will also find implementations of Inner Speech, a new cognitive architecture, inspired by the Rational Speech Acts (RSA) framework.

Getting Started

Please ensure you have anaconda installed. If not, install it using the instructions here: https://www.anaconda.com/download

Run the following bash commands on your terminal:

conda env create -f environment.yaml
conda activate ec

Usage

Training Models

Data: The simulated objects used for training the models are created in the code itself. A random seed is set to ensure the data produced is consistent across execution runs, and to enable comparisons in performance.

1. Open Training.ipynb in a Jupyter environment.
2. Select 'ec' as your python kernel's environment.
3. Click 'Run All', or individually run the relevant cell.

Creating Visualisations from Saved Models and Training Results

1. Open Visualisations.ipynb in a Jupyter environment.
2. Select 'ec' as your python kernel's environment.
3. Click 'Run All', or individually run the relevant cell.

Directory Structure

EC-Submission/
│
├──readme.md
│
├──environment.yaml
│
├──Training.ipynb
│
├──Visualisations.ipynb 
│
├──ModelFiles/
│    ├──exp8.1/
│    ├──exp8.2/
│    ├──...
│    └──exp8.8/
│
└──Images/
     ├──image1.pdf
     ├──...
     └──imageX.pdf