Evolution of Neural Network Structures for Adaptive Mating Call Generation and Mate Location in a Population of Breeding Agents
Tom Cooke, <email@example.com>
Centre for Theoretical and Computational Neuroscience, University of Plymouth
It is now universally accepted within the scientific community that the complex structures observed in the nervous sys- terms of animals and the rich behaviours associated with them did not appear fully formed, but were the result of the action of natural selection over numerous generations [Darwin]. What is the nature of the processes that have enabled complex structures in the brain to emerge through gradual evolutionary change? In order to make some progress towards answering this question, we consider a population of agents with a neural network to control locomotion, mating call production and hearing, and study the evolution of agent behaviour and neural network structures in this model.
A computational model of a population of agents that must use their ability to “sing” and “hear” to locate potential mates in a noisy environment is under development. Agents have a limited amount of energy to expend on singing and movement. The population contains male and female agents; in any season a given female produces offspring with only one male, but a male may mate with many females.
The tension between the need to conserve energy and the need to locate a mate is the primary evolutionary driver in this model. An agent that efficiently produces a mating call that can be clearly distinguished from background noise, or is able to successfully follow mating calls produced by the opposite sex of its own species stands a much better
chance of producing offspring than an agent which does not possess these abilities.
The second key evolutionary driver is sexual selection. When the above-mentioned abilities are sufficiently well developed, competition between male agents becomes stronger and females must become increasingly selective in order to ensure that any male offspring stand a good chance of breeding.
The fitness function of the model allows agents to be compared according to their ability to produce and transduce simple patterns, but also provides opportunities for fitness to be increased by controlling the timing, amplitude, and frequency of the expression of previously evolved patterns.
A system for controlled genetic crossover using marking of structural innovations, developed in the NEAT project [Stanley] is employed in order to allow the protection of network structure through speciation.
The simulation of an early version of the model shows that agents evolve a strategy of a long period of energy conservation followed by a loud burst of singing; this strategy enables them to be heard over background noise. In the evolved population in the full model, the author expects that:
. Agents will evolve the ability to change the loudness, duration, and pitch of bursts in response to the level of noise from other agents.
. Distinct species will emerge, each with its own characteristic song. . Competition between agents will lead to oscillatory dynamics of the singing produced by the agent population as a whole.
Darwin, C. (1859) The Origin of Species. Penguin Classics (1985). Penguin, London. Stanley, K. and Miikkulainen, R. (2002). Evolutionary Computation 10(2): 99-127