Self-Programming: Operationalizing Autonomy

Abstract

Lacking an operational definition of autonomy has considerably weakened the concept's impact in systems engineering. Most current "autonomous" systems are built to operate in conditions more or less fully described a priori, which is insufficient for achieving highly autonomous systems that adapt efficiently to unforeseen situations. In an effort to clarify the nature of autonomy we propose an operational definition of autonomy: a self-programming process. We introduce Ikon Flux, a proto-architecture for self-programming systems and we describe how it meets key requirements for the construction of such systems. Structural Autonomy as Self-ProgrammingWe aim at the construction of machines able to adapt to unforeseen situations in open-ended environments. Adaptation is used here in a strong sense as the ability of a machine not only to maintain but also to improve its utility function and so, in partially specified conditions with limited resources (including time) and knowledge. As a case in point, today's Mars rovers would simply ignore the presence of an alien character waving its tentacles in front of the cameras: observers on Earth would probably see and identify it, but the rover itself would simply not be aware of this extraordinary situation and engineers would have to upload software upgrades to change its mission and plans. In sharp contrast to such engineering, we expect adaptation to be performed automatically, i.e. with no further intervention by programmers after a machine enters service. Our adaptable rover would be fitted with software aiming at discovering facts in a general fashion, that is, not limited to ultra-specific mission schemes. This software would ideally be able to generate new missions and related skills according to the context, within the limitations of its resources -hardware, energy, time horizon, etc. Structural AutonomyThe mainstream approach outlined above consists in the main of sophisticated ways for selecting and tuning hardcoded goals and behaviors for handling situations framed in hard-coded ontologies. Systems designed this way belong to the class of behaviorally autonomous systems [2], and result in fact from the traditional top-down design approach: a machine's operation is fully specified, as is the full range of its operating contexts, and it is the duty of its operator to ensure that the operational conditions always comply to said specification -otherwise the system ceases to function correctly. The point here is that such machines are meant not to change. Adding such change, or adaptation, to the requirements of a machine calls for an orthogonal perspective that addresses change as a desirable and controllable phenomenon. We envision motivations, goals and behaviors as being dynamically (re)constructed by the machine as a result of changes in its internal structure. This perspective -structural autonomy -draws on Varela's work on operationally closed systems [14]: "machine[s] organized (defined as a unity) as a network of processes of production (transforma...

Solving a Problem With or Without a Program

“…In the meanwhile, it tries to build programs for future similar problems. Nivel and Thórisson (2009) argue that an AGI system needs some form of selfprogramming ability. NARS can "self-program", in the sense that • It can form compound operations out of simpler operations.…”

On Attention Mechanisms for AGI Architectures:A Design Proposal

“…The design makes few and fairly high-level requirements for the underlying architecture but is otherwise architectureindependent. The design proposal presented here is just that -a proposal for a design -but the basic principles on which it rests have already been proven in prior architecture implementations (Nivel 2007& 2008, Thórisson 2009a& 2009b. Our work so far has not only resulted in the new attention mechanism presented here but also greatly affected the kinds of architectures we consider to be relevant to AGI research -architecture and attention are co-dependent.…”

Analysis of Types of Self-Improving Software

“…They also list properties of a system which make it purposeful, goal-oriented and selforganizing, particularly: reflectivity -ability to analyze and rewrite its own structure; autonomy -being free from influence by system's original designers (bounded autonomy -is a property of a system with elements which are not subject to selfmodification); endogeny -an autocatalytic ability [30]. Nivel and Thorisson also attempt to operationalize autonomy by the concept of self-programming which they insist has to be done in an experimental way instead of a theoretical way (via proofs of correctness) since it is the only tractable approach [31].…”