بررسی فضای طراحی برای یک سیستم یکپارچه هوشمند

Understanding the trade-offs available in the design space of intelligent systems is a major unaddressed element in the study of Artificial Intelligence. In this paper, we approach this problem in two ways. First, we discuss the development of our integrated robotic system in terms of its trajectory through design space. Second, we demonstrate the practical implications of architectural design decisions by using this system as an experimental platform for comparing behaviourally similar yet architecturally different systems. The results of this show that our system occupies a “sweet spot” in design space in terms of the cost of moving information between processing components.

Intelligent systems (e.g. intelligent service robots) are a product of the many design decisions taken to ensure that the final system meets the requirements necessary to fit in its particular niche [19]. In nature, evolution creates behaviours and bodies that suit an animal’s ecological niche. In the field of intelligent artifacts, choices about the design and implementation of hardware and software may be taken by a designer, or enforced by project or resource constraints. Few, if any, of these choices are truly independent; using a particular solution for one part of the system will constrain the space of solutions available for other parts of the system. For example, the number of degrees of freedom of an effector will restrict the design of the control software and behaviours required to use the effector, and the choice of middleware for software components will restrict the communication patterns that components can use. Understanding the trade-offs available in the design space of intelligent artifacts is a major open issue in the understanding of integrated intelligent systems, and thus AI. In this paper, we focus on the design space of architectures for intelligent robots. We discuss the design of, and the trade-offs created by, an architecture schema for intelligent agents based on a model of shared working memories. Following this we present a novel exploration of the design space of information sharing models for architectures for integrated intelligent systems based on this schema. This exploration uses an intelligent robot as an experimental platform. The robot’s architecture is varied in principled ways to generate quantitative information demonstrating the costs and benefits of the different designs.

From these results we can conclude that a functionally-decomposed n:mn:m CAS instantiation occupies a “sweet spot” in architectural design space with reference to filtering and communication costs. This sweet spot occurs because having too much information shared between components in a system (the n:1n:1 extreme) means that all components incur an overhead associated with filtering out relevant information from the irrelevant information. At the other extreme, when information is not shared by default (the n:nn:n extreme) there are extra communication costs due to duplicated transmissions between pairs of components, and (in CAS-derived systems at least) the “routing” overhead of transmitting information to the correct components (i.e. the filtering performed by working memories rather than components). In this simple example the existence of such a sweet spot, subject to well defined assumptions, could be established mathematically without doing any of these experiments. However, we have shown the possibility of running experiments to test such mathematical derivations, and also to deal with cases where no obvious mathematical analysis is available because of the particular features of an implementation. It is clear that our results are not the end of the story. We have yet to explore n:mn:m instantiations that are not designed along functional lines; it seems sensible to expect them not to perform as well as the n:mn:m system presented here. It is also not clear that n:nn:n instantiations in CAS accurately capture the benefits of a directly connected system, as CAS’s design is tailored to information sharing as a default. This observation leads us to consider an open question: what other appropriate metrics should be considered when evaluating trajectories through design space? In this paper we considered relatively low-level metrics because they could be captured and characterised relatively easily. Other relevant metrics include behavioural measures (e.g. how likely a system is to achieve a goal), expressiveness measures (e.g. how easy it is to encode a particular solution to a problem in an architecture), and meta-level measures (e.g. how easy it is for a designer, or the system itself, to reconfigure the architecture or alter its functionality). It is only when this whole space of possibilities is addressed can we truly start to judge the trade-offs of designing an architecture in a particular way (with reference to a particular task). Even given these shortcomings, the novel experimental methodology presented in this paper points to a route forward for the principled study of integrated intelligent systems in AI. It is our hope that further work along these lines will provide system designers with a body of knowledge about the choices and trade-offs available in architectural design space, allowing them to build systems that satisfy their requirements in an informed and principled manner.