Abstract A central challenge in artificial life concerns how agents engage with open-ended spaces of action possibilities, or affordances, whose full extent cannot be specified in advance. Classical arguments suggest that artificial systems are limited in this regard, as they operate within predefined state and action spaces, unlike biological organisms that continually generate new affordances. This paper revisits this limitation in light of recent developments in artificial intelligence. We argue that contemporary artificial agents increasingly operate within environments whose interfaces and infrastructures help structure the relations through which action possibilities become accessible. To capture this phenomenon, we introduce the concept of scaffolded affordances: actionable possibilities encoded in interfaces connecting agents to distributed resources. We develop a taxonomy of affordance engagement—possession, access, and orchestration—and a minimal formal model in which affordances arise from agent–environment interaction. Within this framework, agents need not enumerate all possible actions internally, but may instead navigate and coordinate affordance relations scaffolded by their environments. We argue that such systems operate within affordance ecologies, where the space of actionable possibilities is shaped by evolving environments. This perspective suggests that artificial systems may approximate aspects of open-ended agency through the expansion and reconfiguration of their environments, providing a bridge between ecological theories of affordances and artificial life research on open-ended behaviour.
Abstract Parental care is prevalent, especially among mammals. A long-lasting research question lies in whether Post-Maturity Parental Investment (PMPI) has any behavioral or physical population effect. However, it is ethically inconceivable to directly experiment on real mammals in a sizable manner. Therefore, AI agent populations with and without PMPI are proposed to examine the population-scale consequences under different scenarios. Specifically, both multi-agent reinforcement learning (MARL) experiments and multi-agent simulation (MAS) experiments are proposed for investigating the PMPI effects on population dynamics. MARL tests the behavioral aspect (learning) while MAS tests the physical aspect (survival). Based on millions of simulated steps, the results suggest that PMPI secures lineage successions and leads to family-dominated formations at individual level. However, at population level, PMPI is negatively correlated with population size and total wealth, while being positively correlated with wealth inequality (as measured by the Gini index). Both experiments suggest that extended parental investment (e.g., PMPI) imposes systemic constraints on population dynamics. On the other hand, those simulated results can be corroborated by real world demographic data from global authoritative sources. The MAS source code, executable, and web version is available for scientific reproducibility at https://www.cs.toronto.edu/~wkc/ParentalCaresOnline.html.
The Italian-Norwegian scientist Nils Aall Barricelli conducted pioneering work on the evolution of digital organisms in the 1950s and 1960s. His paper Esempi Numerici di Processi di Evoluzione (Numerical Examples of Evolutionary Processes), published in the Italian journal Methodos in 1954, is, to the best of our knowledge, the first publication describing experiments in computational evolution and artificial life. Had the paper been more widely known when the field of Artificial Life coalesced as a recognised discipline in the 1980s, there is little doubt that it would have greatly influenced early research directions and established Barricelli as a founding father of the discipline. However, despite the historical significance of the paper, it has until now remained a fairly obscure publication—Methodos ceased publication in 1964, and copies of the paper are hard to find online. In publishing an English translation of the paper in this issue of the Artificial Life journal, we aim to provide a much more widely accessible version of the work. In this accompanying commentary, we provide some scientific and biographical context for the paper, discuss how it relates to Barricelli’s earlier and later publications, comment on the process of translating the work and issues that arose in the process of doing so, and reflect on the legacy and significance of the work today.
Developmental Graph Cellular Automata (DGCA) are a novel model of morphogenesis, capable of growing directed graphs from single-node seeds. In this paper, we extend the DGCA model to exhibit stochastic growth, and show that both deterministic and stochastic DGCAs can be trained to grow reservoirs. Reservoirs are grown with two types of targets: task-driven (using the NARMA family of tasks) and task-independent (using reservoir metrics). Results show that deterministic DGCAs are able to grow a variety of specialised, life-like structures capable of solving benchmark tasks, outperforming 'typical' reservoirs on the same task. Stochastic DGCAs grow robust, regular structures that outperform reservoirs grown deterministically. Overall, these lay the foundation for the development of DGCA systems that produce plastic reservoirs and for modelling functional, adaptive morphogenesis.
