\documentclass[letterpaper]{article}% DO NOT CHANGE THIS
\def\year{2021}\relax
%File: formatting-instructions-latex-2021.tex
%release 2021.1
\documentclass[letterpaper]{article}% DO NOT CHANGE THIS
\usepackage{color}
\usepackage{aaai21}% DO NOT CHANGE THIS
\usepackage{times}% DO NOT CHANGE THIS
\usepackage{helvet}% DO NOT CHANGE THIS
...
...
@@ -158,8 +161,18 @@
\maketitle
\begin{abstract}
Recent advance on symbolic language in neural network based multi-agent systems have shown great progress in compositionality, which is taken as a distinguished feature of human language different from animal language. However, these efforts only explored environmental pressures, without realizing the importance of characterization capacity of agents.
In this work, we explore the relationship between the characterization capacity of agents and the compositionality of symbolic languages. By both proving with mutual information theory and verifying with extensive experiments, we made the counter-intuitive conclusion that symbolic languages with higher compositionality require lower characterization capacity of agents and are easier-to-teach.
Recent advances on symbolic language in neural network based multi-agent systems
have made great progress in compositionality, which is taken as a key
feature distinguishing human language from animal language. However, these efforts
only explored environmental pressures, without realizing the importance of
characterization capacity of agents.
In this work, we explore the relationship between the characterization capacity
of agents and the compositionality of symbolic languages. By proving with
mutual information theory and verifying with extensive experiments, we made the
counter-intuitive conclusion that symbolic languages with higher
compositionality require lower characterization capacity of agents and are
