Content area
Abstract
Interactions between multiple competing and cooperating populations can drive the evolution of complexity. In this thesis, we will demonstrate that specific topologies of interactions are particularly successful at developing complexity, and will apply these lessons to a class of multi-agent deep learning algorithms, thereby demonstrating the parallels between the fields of artificial life and deep learning. Our analysis will cover multiple different evolutionary models, focusing on the paradigm of matrix games, which serve as an effective testbed for hypotheses about simulated evolution.
In seeking to measure the evolution of complexity, the problem of defining a clear and computable metric of complexity is critical. We will present an overview of complexity definitions and formulas from the literature, and show how techniques can be adopted from automata theory and information theory to create meaningful metrics for specific classes of artificial organisms.
We will leverage these metrics to present experiments which show that topologies which mix cooperative and competitive interactions are able to generate faster complexity growth, and sustain that growth over longer time scales than pure competitive arms races or pure cooperative symbiosis. We will present analysis of the roadblocks that stall complexity growth in unsuccessful topologies, and the manner in which successful topologies overcome these obstacles.
Many of the phenomena which have been studied in the field of artificial life are being rediscovered, and re-examined in the field of deep learning (DL), especially in models with multiple interacting DL agents. The same obstacles which prevent successful evolution of complex forms also stymie these deep models, and the same tricks which have been developed by artificial life researchers to overcome them are now being developed by DL researchers. We will demonstrate how lessons from our models can be translated to the realm of DL to improve model performance.
