Included here is a general discussion of a set of problems that can be taken together as being aspects of what we might call "self-evolution".
Evolution patiently accumulates adaptations, but self-evolution, which intentionally remakes itself in response to the unexpected, is trickier. Both evolution and self-evolution are based on selective feedback loops. In the case of self-evolution, the feedback loop selects for the ability to evolve. Over time, a successful self-evolving system will be able to evolve more effectively. When a system selects for structures that enable it to evolve more effectively, a meta-selective pilot emerges within that system to direct the system's own development.
Self-evolution adds another core principle. When a system selects for its own ability to evolve, a non-mechanical pilot emerges. The pilot can serve as a handle or interface for the system. Self-evolving pilots enable a range of applications that involve coupling complex systems to each other. The coupling makes it possible to create meaningful interactions between the systems, without depending on or referring to an explicit, analytical representation of either system.
DNA-based computation is essentially a massively parallel form of digital computation that uses base pairs to represent data. Because DNA can be used to perform many computations at once, it is useful for solving problems that a programmable electronic computer would take far too long to solve. Research may eventually show that self-evolving processes are at least indirectly involved in DNA computation.
Evolution depends on open protocols that allow otherwise independent structures to interoperate. Open protocols form a stage upon which further development can unfold, making development possible without necessarily causing or directing it. As an example, although DNA is often thought of as a basis for encoding and transmitting information, perhaps a more fruitful concept is that DNA is a protocol around which complex developmental activities spontaneously organize.
One interesting example, named for the mathematician Joseph Plateau, is the use of soap bubbles for finding solutions to minimal surface problems, or for an NP-complete problem such as the traveling salesman problem, that are resistant to computational solutions.
Software can't write itself. An outside agent, the software developer, has to supply the code. A larger code structure, based for example on templates, could choose from the syntactically possible code in each situation, but then an outside agent would have to author that larger structure, and that agent would indirectly also be the author of what the template module produces. The regression is endless. Von Neumann showed that self-constructing machines are possible, but self-construction is dimensionally distinct from self-design. Software is essentially mechanical, with all possible states specified in advance, whereas evolution is surprise.
Software that selects code fragments based on the results produced by those fragments appears to be creative, but careful examination of categories shows that such creativity is really an automated unfolding of the programmer's own intentions. Genetic algorithms and "artificial life" are interesting from a software engineering perspective, but are misleading if applied to questions of the nature of life or of mind.
No, the premises are categorically different. AI seeks to implement intelligence through computation. Even biologically inspired approaches, such as connectionism or artificial neural nets, are computational at root. Although computers are certainly useful tools for researching and developing self-evolving systems, self-evolution itself is not based on computation.
The usual view of evolution involves a selection mechanism operating on a source of random variation. That view implies that any source of randomness would work, because the "magic" lies in the selection mechanism. However, in the rich context of biological systems, events that are supposedly random are deeply implicated with the system as a whole. What appears to be colorless, featureless randomness is actually pregnant with potential structure. Biological evolution is an unfolding or realization of that potentiality. The pseudo-randomness generated by computer software is therefore categorically unsuitable. AL is interesting as a software design methodology, but it isn't evolutionary in the biological sense.
The difficulty cuts deeper than the debate over digital technology versus analog. Self-evolution requires that systems be non-deterministic and thermodynamically open. Although some software-controlled systems exchange matter, energy, and information with their environment, software itself consists strictly of predetermined rules.
A ubiquitous, global digital network will soon allow devices to share services with each other seamlessly. Many new applications for networked digital devices will be developed. The result will be a roiling, ebullient ocean of data that will be too vast and dynamic to interface easily with self-evolving systems. That makes it unlikely that self-evolving systems will ever come remotely close to displacing digital computers. Certain hybrid applications might eventually be possible, however, if the evolutionary processes within self-evolving systems can be made rapid enough.
Because they are dynamical and recurrent, self-evolving systems are difficult to model explicitly. Almost any analytical approach will model the systemís behavior on the basis of unacceptably transient or partial views. Working with the system as a whole requires an approach that abandons explicit analysis altogether and instead uses complex systems as tools for studying other complex systems, without modeling any of the systems directly. This black-box-over-black-box approach never aims for a complete understanding. Rather, it involves developing techniques for specific results. Self-evolving systems themselves can serve as useful instruments in this context.
Part of the challenge is to be honest and modest regarding what can be known about nature. Classical science and engineering succeed by idealizing what would otherwise be inscrutably complex aspects of nature. Restricting the domain of analysis gives the illusion of certain knowledge. However, when dealing with complexity such as that involved in protein folding, the messiness itself is part of the essence of the problem and cannot be systematically dismissed. Non-representational analysis can be quantitative, although it never claims to yield definite or complete solutions.
An important new frontier in science will involve the use of complex systems to explore and manipulate other complex systems. One area where this approach will be used will be in developing proteins and other biomolecules for specific applications. Self-evolving pilots will be used where complexity makes an analytical approach unfeasible. The self-evolving approach will be crucial in genomics and proteomics. Even the most powerful parallel processing computers will be overwhelmed by the full complexity of protein folding and of the precise mapping of genome to phenotype.
Self-evolution is similar to autopoiesis in that both involve coordination through self-referential interactions. A self-evolving system can have many of the properties of autopoietic systems. It may have operational closure, for example, or it may involve systems that co-evolve. Autopoiesis emphasizes the self-maintaining nature of a system. Self-maintenance is necessary but not sufficient for self-evolution, however.
Perhaps our inner sense that the self is real and is the seat of conscious experience (Zen Buddhists explore this inner sense through careful introspection) is produced by an ensemble of self-evolving pilots in the brain. Perhaps self-evolving processes in the brain are also the basis of our universal human moral sense.
Michael Webb, 2000
[ home ]