The emergence of modular agents in complex environments
Supervisor: Dr. Juan Neirotti

Background

There is no doubt that the human brain is one of the most complex known biological systems; it is a modular system characterised, among other things, by the parallel processing of information and multitask capabilities. 

In a very general and abstract sense, modular systems can be defined as systems constructed from structurally and functionally distinct parts. In the current proposal we will consider the following  types of modularity [1]:

	a) Architectural modularity: the system is formed by several subunits or architectural modules, that are mainly detached from one another and may perform different tasks. Architectural brain modules are, for instance, the cerebellum and the neocortex.
	b) Functional modularity: characterised by differences in the task performed, which does not seem to be based on underlying differences in structure or computational process. Different architectural modules may work together to cope with a given task, acting as a functional module. If the task is finished or changed, the functional module may cease to exist. Functional  modules can be put in evidence by functional MRI studies.

We will apply artificial life simulations to  the study of populations of learning agents that perform several tasks simultaneously. For the simulations we will consider networks composed of interlinked units, with an inherited architecture. Each unit possesses a program that represents its learning algorithm.  Evolution will be mimicked by Genetic Programming (GP) [2] operating over the learning programs, and by Genetic Duplications [3], modifying the agent's architecture. In this form, evolution operates over a wider space of possible variations, modifying both the hardware (architecture) and the software (program) of the agent.  

The study of the behavioural response of agents, simulated by these techniques, and interacting with different environments, will help us understand whether modularity is a persistent property, or a consequence of the complexity of the environment. To understand this we will study the correlation between the measure of complexity of the environment (as defined in [4]) and the measure of modularity. 

Aims and Objectives

The main thrust of this project is to gain insight into the origin and functionality of  modules in an agent based learning system that lead to an improvement in the performance several tasks simultaneously. We would like to identify the mechanisms that drive the evolutionary process to favour the emergence of such modules at different stages of development. More precisely, we aim to:

	a) Identify the characteristics of a suitable environment that drives the emergence of specialised modules in evolving learning agents.
	b) Devise an improved simulation technique based on GP to make populations of learning agents evolve in the aforementioned environment, and identify the chronological order, if any, of the emergence of the relevant modules. 

[1] J. Bryson & L. A. Stein, Modularity and specialised learning: mapping between agent architectures and brain organisation, in Emergent neural computational architectures based on neuroscience: towards neuroscience-inspired computing, Springer-Verlag, New York, 2001.
[2] J. R. Koza, Genetic Programmin}, MIT Press, Cambridge, MS, 1992.
[3] S. Ohno Evolution by Gene Duplication. Springer, Berlin, 1970.
[4] L. Franco & M. Anthony, IEEE Trans. N. N.  17, 578 (2006).