My PhD Thesis: Evolutionary Computational Intelligence for Multi-Agent Simulations
You can download a pdf version of my thesis here.
Abstract
Keywords: Computational Intelligence; Multi-Agent Simulations; Evolutionary Computation; Genetic Programming; Complex Systems; Artificial Life; Emergence of Group-Behavior; Bloat Control.
The growing interest in multi-agent simulations, influenced by the advances in fields like the sciences of complexity and artificial life is related to a modern direction in computational intelligence research. Instead of building isolated artificial intelligence systems from the top-down, this new approach attempts to design systems where a population of agents and the environment interact and adaptation processes take place.
We present a novel evolutionary platform to tackle the problem of evolving computational intelligence in multi-agent simulations. It consists of an artificial brain model, called the gridbrain, a simulation embedded evolutionary algorithm (SEEA) and a software tool, LabLOVE.
The gridbrain model defines agent brains as heterogeneous networks of computational building blocks. A multi-layer approach allows gridbrains to process variable-sized information from several sensory channels. Computational building blocks allow for the use of base functionalities close to the underlying architecture of the digital computer. Evolutionary operators were devised to permit the adpative complexification of gridbrains.
The SEEA algorithm enables the embedding of evolutionary processes in a continuous multi-agent simulation in a non-intrusive way. Co-evolution of multiple species is possible. Two bio-inspired extensions to the base algorithm are proposed, with the goal of promoting the emergence of cooperative behaviors.
The LabLOVE tool provides an object model where simulation scenarios are defined by way of local interactions. The representation of simulation object features as symbols mitigates the need for pre-defined agent sensory and action interfaces. This increases the freedom of evolutionary processes to generate diversified behaviors.
Experimental results are presented, where our models are validated. The role of the several genetic operators and evolutionary parameters is analyzed and discussed. Insights are gained, like the role of our recombination operator in bloat control or the importance of neutral search. In scenarios that require cooperation, we demonstrate the emergence of synchronization behaviors that would be difficult to achieve under conventional approaches. Kin selection and group selection based approaches are compared. In a scenario where two species are in competition, we demonstrated the emergence of specialization niches without the need for geographical isolation.