Evolving

Self-organisation

Recent dramatic advances in the problem-solving capabilities and scale of Artificial Intelligence (AI) systems have enabled their successful application in challenging real-world scientific and engineering problems (Abramson et al 2024, Lam et al 2023). Yet these systems remain brittle to small disturbances and adversarial attacks (Su et al 2019, Cully 2014), lack human-level generalisation capabilities (Chollet 2019), and require alarming amounts of human, energy and financial resources (Strubel et al 2019).

Biological systems, on the other hand, seem to have largely solved many of these issues. They are capable of developing into complex organisms from a few cells and regenerating limbs through highly energy-efficient processes shaped by evolution. They do so through self-organisation: collectives of simple components interact locally with each other to give rise to macroscopic properties in the absence of centralised control (Camazine, 2001). This ability to self-organise renders organisms adaptive to their environments and robust to unexpected failures, as the redundancy built in the collective enables repurposing components, crucially, by leveraging the same self-organisation process that created the system in the first place.

Self-organisation lies at the core of many computational systems that exhibit properties such as robustness, adaptability, scalability and open-ended dynamics. Some examples are Cellular Automata (Von Neumann 1966), and their more recent counterparts such as Neural Cellular Automata (Mordvintsev et al 2020) and Lenia (Chan, 2019), reaction-diffusion systems (Turing 1992, Mordvintsev 2021), and particle systems (Reynolds 1987, Mordvintsev) . Examples from neuroevolution are indirect encodings of neural networks inspired from morphogenesis such as cellular encodings (Gruau 1992), HyperNEAT (Stanley et al 2009), Hypernetworks (Ha 2016), HyperNCA (Najarro et al 2022) and Neural Developmental Programs (Najarro et al 2023, Nisioti et al 2024), showing improved robustness and generalisation. Aside biological evolution, cultural evolution can also lead to open-ended systems, such as human culture, aspects of which have been captured by computational models of social dynamics, as with Schelling's model (Schelling, 1978), Spatial Social Dilemmas (Nowak and May, 1992) and, more recently, groups of Large Language Models (Nisioti el al, 2024).

Guiding self-organising systems through evolution is a long-standing and promising practise, yet the inherent complexity of the dynamics of these systems complicates their scaling to domains where gradient-based methods or simpler models excel (Risi 2021). If we view self-organising systems as genotype to phenotype mappings, we can leverage techniques developed in the evolutionary optimization community to understand how they alter evolutionary dynamics and guide them better.

The reverse is also possible: evolution can emerge as an inherent property of a self-organising system allowing us to study questions about the origin of life. Investigating under which conditions they appear, and the particular emergent evolutionary behaviours in these systems could afford insights applicable to existing artificial evolutionary approaches, or even directly provide an evolutionary substrate for learning tasks and achieving open-endedness. Early work in this direction (Ray 1992, Agüera y Arcas et al 2024, Fontana 1990, Adami et al 1994, Rasmussen et al 1991) has demonstrated emergent evolution in several computational substrates.