In principle, a self-organising system cannot be constructed, since its organisation and behaviour cannot be prescribed and created by an external source. It emerges autonomously in certain conditions (which cannot be prescribed either). The task of the researcher is to investigate in what kind of systems and under what kind of conditions self-organisation emerges.

A self–organizing system acts autonomously, as if the interconnecting components had a single mind. And as these components spontaneously march to the beat of their own drummer, they organize, adapt, and evolve toward a greater complexity than one would ever expect by just looking at the parts by themselves.

The first attempts to consider the behavior of so-called "random neural nets" in a systematic way have led to a series of problems concerned with relations between the "structure" and the "function" of such nets. The "structure" of a random net is not a clearly defined topological manifold such as could be used to describe a circuit with explicitly given connections. In a random neural net, one does not speak of "this" neuron synapsing on "that" one, but rather in terms of tendencies and probabilities associated with points or regions in the net.

The computer scientists Jeff Clune, Jean-Baptiste Mouret, and Hod Lipson did what computer scientists do: they designed computer simulations.23 They used well-studied networks that had sensory inputs and produced outputs. What those outputs were determined how well the network performed when faced with environmental problems. They simulated twenty-five thousand generations of evolution, programming in a direct selection pressure to either maximize performance alone or maximize performance and minimize connection costs. And voilà! Once wiring-cost-minimization was added, in both changing and unchanging environments, modules immediately began to appear, whereas without the stipulation of minimizing costs, they didn’t. And when the three looked at the highest-performing networks that evolved, those networks were modular. Among that group, they found that the lower the costs were, the greater the modularity that resulted. These networks also evolved much quicker — in markedly fewer generations — whether in stable or changing environments. These simulation experiments provide strong evidence that selection pressures to maximize network performance and minimize connection costs will yield networks that are significantly more modular and more evolvable.

The main theme of this report is a particular facet of the general problem of pre-organization in self-organizing systems, namely, the theory and circuitry of information processing networks. One may consider these networks as a special type of parallel computation channels which extract from the set of all possible inputs a particular subset which is defined by the internal structure of the network. The advantage of such operationally deterministic networks in connection with adaptive systems is the obvious reduction in channel capacity of the adaptors, if it is possible to predetermine classes of inputs which are supposed to be meaningful for those interacting with the automaton