All evolutionary systems, natural or artificial, are built from essentially the same three elements: variation, inheritance and selection. What then distinguishes the process of biological evolution by natural selection, which produces such impressive outcomes, from the relatively underwhelming results of artificial digital evolution? We focus on one aspect of this: the emergence of simple replicators from a lower-level foundation in an artificial chemistry. Previous work has either supplied a handbuilt basic (or shortcut) replicator for evolution to work upon, or has provided direct support for replication in the chemistry itself.

Our first research question concerns the relationship between heredity and selective pressure in a theoretical replicator. We construct a simple model of generalized evolution that shows complex and non-obvious emergent behaviour. We show by simulation that inheritance in this model is a target of evolution, and that it evolves under a range of conditions. The degree of inheritance is related to the predictability of environmental change, and the degree of inheritance is tuned by evolution to balance fitness and robustness. Fitness is maximized in unchanging environments where there is little penalty to reduced diversity, while a more diverse population is maintained in changing environments to provide robustness to environmental change. This balance emerges unprogrammed from the underlying model.

Our second research question regards the practicality of realising replication in an artificial system. The investigation is founded on ToyWorld, a highly-modular artificial chemistry that allows us to explore the effect of different combinations of modules, such as for reactant or product selection, upon replicator formation. Our underlying hypothesis is that replicators can form from sequences of linked reaction cycles, where the stochiometry of the sequence is necessarily greater than one for replication.

We first test the influence of two strategies to select reactants for a reaction, and two other strategies for selecting the resulting products post-reaction from the alternative productsets. Our first reactant selection strategy is to choose reactants with equal-probability from the set of possible reactants without consideration of position; our second strategy is spatial, where reactants are chosen if they are spatially co-located. For product selection, the first strategy is again based on an equal-probability choice from the alternative product sets, while the second biases the choice towards the product set with the greatest energy return, or least energy input (a “least-energy” strategy). Of the four possible combinations of reactant and product selection strategies, the combination of spatial reactant selection and least-energy product selection strategies maximizes the number and cycle-length of the resulting reaction cycles.

Next we search the ToyWorld reaction network for evidence of exact multipliers: repeated increasing sequences of the same type of reaction cycle, connected by one or more shared reactants and products. We develop an algorithm for the detection of multipliers in a reaction network, and using that, we find that multipliers do form in ToyWorld, but at low rates, and without great longevity. They occur as a result of Product and Reactant selection strategies, and not by chance alone.

Finally, we examine the reaction network for variable replicators: multipliers that can take any of a set of structural states that appear equivalent under selection. We extend our earlier model of external environmental change, and search for variable replicators in the resulting reaction network. Our results are inconclusive. Candidate variable replicators emerge, but each is endemic to a single run.