Academic journal article Cosmos and History: The Journal of Natural and Social Philosophy

Universal Principles of Intelligent System Design

Academic journal article Cosmos and History: The Journal of Natural and Social Philosophy

Universal Principles of Intelligent System Design

Article excerpt

I. INTRODUCTION

We live in the era of yet another technological revolution. If the previous one was associated with the development of energy conversion tools to produce work, harnessing increasingly vast resources over the course of history, the world nowadays is being rapidly transformed by the information processing tools. Starting with arithmetic devices that support conditional clauses and store intermediate results, we have arrived at machines emulating major functions that have before typically been attributed only to living beings or specifically to humans--learning, knowledge management, pattern recognition and control. The source of hopes and fears these days is being called <>, potentially dramatically enhancing the life conditions for humans and at the same time outcompeting us in the labor market. There is also quite an amount of speculation and research concerning theoretically possible super-intelligences posing existential risk for our species, and whether it is possible to align their interests with ours. Despite the massive effort in the field, the big question remains standing in the background: What is intelligence?

There have been numerous attempts at defining intelligence. In those proposals, the concept of intelligence often involves at least two distinct notions--one concerning the performance of an entity in a given fixed environment, and another one concerning its ability to learn new environments. The emphasis depends on the preferences of individual authors. It is also notable that these definitional inherently teleological, describing the intelligent entity as having a goal or receiving rewards. The most elegant and general definition [1] is as well set in the framework of reinforcement learning and involves uncomputable universal probability distribution which precludes its practical use.

The notion of intelligence in a system is strongly associated with its information processing architecture [2,3]. Luckily, information is a well-defined property of a signal. Moreover, since the advent of information era, a number of connections were established, relating information theory to statistical physics and thermodynamics. By now the energy-information duality is fundamental to our understanding of organization in complex systems and emergence of functionality. Intelligence is a functional property of systems, therefore it is supposed to have complementary energetic and informational descriptions. Using the connection between information theory and physics provided by computational mechanics, the question of intelligence becomes physically tractable.

In this paper we propose a universal definition of intelligence for a physical system and discuss its implications on design of intelligent systems.

II. GUIDING PRINCIPLES

We want our definition of intelligence to have certain properties:

1. Physical: intelligence is well-defined in terms of physical observables of the system.

2. Universal: the definition should be applicable to any realizable physical system without referring to anthropic metaphors (e.g. goals or rewards) or particularities of the realization.

3. Observable: intelligence of an entity should be in principle observable by the third parties.

4. Practical: intelligence of an entity should be determinable in finite time using only the observation of its behavior. Note that the word behavior here is used in a more general sense than its traditional use in psychology, and refers to the path of system in its configuration space as it evolves in spacetime, or some coarse-grained version of it that can be obtained from measurement.

5. Partial order: we want to be able to compare intelligences of different systems, grading them as more or less intelligent.

6. Causal: intelligence of an entity should depend only on the regions of spacetime causally accessible from it. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.