PhD thesis available: Autonomous Formation of Concepts and Communication

[Edwin de Jong]

Dear Connectionists,

this is to announce the availability of my PhD thesis for download.

URL: http://arti.vub.ac.be/~edwin/thesis

Title: Autonomous Formation of Concepts and Communication

Best regards,

Edwin de Jong
__

Keywords: concept formation, evaluative feedback, reinforcement 
learning, situation concepts, development of communication, dynamical 
systems.

Abstract

Autonomous agents receive sensor input from the environment, select 
actions, and may receive evaluative feedback on their behavior. 
Multiple agents that are present in the same environment can benefit 
from using communication to overcome uncertainty in their information 
about the environment. The research in the thesis addresses the 
question of how concepts about the environment can be formed, and how 
a system of communication can develop that allows agents to exchange 
information about their environment. Instead of assuming that concepts 
are already available or based on universal primitives, in the 
approach followed here they are formed in response to interaction with 
the environment. Evaluative feedback can play an important role in 
this. This distinguishes the approach from concept learning, which 
requires supervised feedback.

A particular type of concepts is described, called situation concepts. 
Situation concepts consist of features in the history of interaction 
between the agent and its environment, and predict some aspect of the 
future evolution of the state of the environment, possibly conditioned  
on the actions the agent may take. Several existing methods, 
particularly from the field of reinforcement learning, can be viewed 
as constructing a form of situation concepts, and a particular method 
for constructing a specific type of situation concepts is described. 
Since situation concepts convey information about the environment, they 
are especially suited for use in communication.

The development of communication consists of the formation of 
associations between words and the concepts formed by individual 
agents, and is viewed as the behavior of a dynamical system. The 
variables of this system are the strengths of the associations between 
words and the concepts of all agents in a population. An algorithm for 
association formation for individual agents is described that leads to 
a common system of communication.

A deterministic version of the system is shown mathematically and 
demonstrated experimentally to have point attractors that correspond to 
perfect communication. The stochastic system system has points in its 
phase space that play a similar role, and is preferable since it avoids 
certain deadlock situations. This finding is confirmed by an 
investigation of the relation between the amount of stochasticity and 
the quality of communication. The research contributes to the view of 
the development of communication as the behavior of a dynamical system. 
Finally, systematic measures are provided for the quality of conceptual 
systems and communication systems that can be used when the subject of 
communication, called referent, is known.