The central goal of this thesis is to provide additional criteria towards implementing open-ended evolution in an artificial system. Methods inspired by biological evolution are frequently applied to generate autonomous agents too complex to design by hand. Despite substantial progress in the area of evolutionary computation, additional efforts are needed to identify a coherent set of requirements for a system capable of exhibiting open-ended evolutionary dynamics. The thesis provides an extensive discussion of existing models and of the major considerations for designing a computational model of evolution by natural selection. Thus, the work in this thesis constitutes a further step towards determining the requirements for such a system and introduces a concrete implementation of an artificial evolution system to evaluate the developed suggestions. The proposed system improves upon existing models with respect to easy interpretability of agent behaviour, high structural freedom, and a low-level sensor and effector model to allow numerous long-term evolutionary gradients. In a series of experiments, the evolutionary dynamics of the system are examined against the set objectives and, where appropriate, compared with existing systems. Typical agent behaviours are introduced to convey a general overview of the system dynamics. These behaviours are related to properties of the respective agent populations and their evolved morphologies. It is shown that an intuitive classification of observed behaviours coincides with a more formal classification based on morphology. The evolutionary dynamics of the system are evaluated and shown to be unbounded according to the classification provided by Bedau and Packard’s measures of evolutionary activity. Further, it is analysed how observed behavioural complexity relates to the complexity of the agent-side mechanisms subserving these behaviours. It is shown that for the concrete definition of complexity applied, the average complexity continually increases for extended periods of evolutionary time. In combination, these two findings show how the observed behaviours are the result of an ongoing and lasting adaptive evolutionary process as opposed to being artifacts of the seeding process. Finally, the effect of variation in the system on the diversity of evolved behaviour is investigated. It is shown that coupling individual survival and reproductive success can restrict the available evolutionary trajectories in more than the trivial sense of removing another dimension, and conversely, decoupling individual survival from reproductive success can increase the number of evolutionary trajectories. The effect of different reproductive mechanisms is contrasted with that of variation in environmental conditions. The diversity of evolved strategies turns out to be sensitive to the reproductive mechanism while being remarkably robust to the variation of environmental conditions. These findings emphasize the importance of being explicit about the abstractions and assumptions underlying an artificial evolution system, particularly if the system is intended to model aspects of biological evolution.

Artificial intelligence has traditionally been approached through centralized architectures and optimization of specific metrics on large datasets. However, the frontiers of fields spanning cognitive science, biology, physics, and computer science suggest that intelligence is better understood as a multi-scale, decentralized, emergent phenomenon. As such, scaling up approaches that mirror the natural world may be one of the next big advances in AI. This thesis presents Coralai, a framework for efficiently simulating the emergence of diverse artificial life ecosystems integrated with modular physics. The key innovations of Coralai include: 1) Hosting diverse Neural Cellular Automata organisms in the same simulation that can interact and evolve; 2) Allowing user-defined physics and weather that organisms adapt to and can utilize to enact environmental changes; 3) Hardware-acceleration using Taichi, PyTorch, and HyperNEAT, enabling interactive evolution of ecosystems with 500k evolved parameters on a grid of 1m+ 16-channel physics-governed cells, all in real-time on a laptop. Initial experiments with Coralai demonstrate the emergence of diverse ecosystems of organisms that employ a variety of strategies to compete for resources in dynamic environments. Key observations include competing mobile and sessile organisms, organisms that exploit environmental niches like dense energy sources, and cyclic dynamics of greedy dominance out-competed by resilience.