Doelen van de AI

Inleiding

Marvin Minsky geeft in Semantic Information Processing (MIT Press '68) een van de meest bekende definities van A.I.:

the science of making machines do things that would require intelligence if done by men.

Een klassieker: de Turing test

Dennet's reactie op Hofstadter's Coffeehouse Conversation over de Turing test (in: The Mind's I).

Marvin Minsky vraagt zich in The Society of Mind het nut af van definities voor intelligentie:

Herb Simon beschrijft in The Sciences of the Artificial een groep systemen (the artificial) en hun relatie tot de wereld zoals die zonder toedoen van de mens is (the natural), en

Patrick Henry Winston deelt in zijn boek Artificial Intelligence de doelen van de A.I. in in 2 soorten: computers slimmer maken om (A) intelligentie beter te begrijpen, en (B) handigere computers te verkrijgen. Verder de mening van Alan Newell in Unified Theories of Cognition.

Begin van: Luc Steels. 1996. The origins of intelligence. In: Proceedings of the Carlo Erba Foundation conference on Artificial Life. Fondazione Carlo Erba. Milano.

The origins of intelligence

Luc Steels

Introduction

The reason is partly that for many people no such theory is needed. Mind is eternal, they say, belonging to a Platonic universe. Seeking an explanation for its origins is therefore absurd. Such a Platonic view is still common today with mathematicians like Penrose [15]. However it is not a scientific explanation. It is similar to the earlier view that the origin of the universe needs no explanation because it has always been there and will always be there, or that all the different species, including Man, were created in a few days by an omniscient being. If we want a scientific theory of the origins of intelligence, we must close the gap between the basic laws of physics and biology and theories of intelligence. Right now the gap is enormous and it can only be closed by working from both sides.

This paper raises a few issues and provides some directions and experimental approaches for addressing the question of the origins of intelligence. No clear definite answer can be given yet, although a way can be pointed out. The first section defines intelligence as a continuum with current biological views of living systems. It is only by having such a definition that we can hope to pinpoint precisely where intelligent systems outgrow living systems.
 

Defining Intelligence

This section proposes an alternative definition of intelligence which seeks to establish a continuum with life. It first identifies the class of evolving complex adaptive systems, then identifies progressively more complex instances from chemical systems to living systems, and then to intelligent systems.

Evolving Complex Adaptive Systems

• Self-maintenance: Self-maintenance means that the system is actively establishing itself. To avoid annihilation due to increased entropy, the system needs to constantly rebuild itself by drawing materials from the environment and establish a boundary between itself and the rest of the environment. Maturana and Varela have called this process autopoiesis [11]
• Adaptivity: The system is not only capable to maintain its own internal equilibrium for a constant environment, but also adapts when there are (small scale) changes to the environment in order to enhance its chances of further existence.
• Information preservation: The information defining the system is capable to be perserved so that the system does not depend on the continued existence of its components to survive. It is the role of the components that keeps the whole system together and if the various roles and their interrelations are preserved the whole system is preserved.
• Spontaneous increase in complexity: The most remarkable aspect is that the system is able to increase its own internal complexity. This could mean that there are increasingly more parts, more complex relations between parts, more complex behaviors of the parts, etc. Moreover often instances of the same system come together to form a larger whole that operates as a single system at a higher level.

1. Autocatalytic chemical reactions (uncoded life)

The various properties of evolving complex adaptive systems can already be seen in certain types of chemical reactions which are known as pre-life or uncoded life systems [8]:

• The reactions achieve self-maintenance by being autocatalytic. The substances to start the beginning of the reaction are regenerated, often after a long cycle and in larger quantities, so that the whole reaction chain can start again and proliferate. In some cases it is possible to show that boundaries form themselves [9].
• These reactions can be shown to be adaptive to changes in the environment. For example, the rate may slow down when temperature conditions change or when materials are less abundantly present. In some cases there are conditional pathways depending on the conditions in the environment.
• Autocatalytic reaction networks preserve information by making copies of themselves (with potential errors). Such reactions have been shown in the laboratory.
• Autocatalytic reactions have recently been shown to be able to undergo evolution by natural selection, known in this case as molecular evolution. It is enough that there is a reaction that is autocatalytic and that variations occur in replication. When the environment (in this case the other chemicals present) provide selectionist pressures, then there is an evolution towards more complex molecules or reaction pathways that are capable to cope better with the selection pressures.

Living systems clearly have all the properties of evolving complex adaptive systems. They most probably originated out of autocatalytic chemical reaction networks but achieve the characteristics of evolving complex adaptive systems differently:

• The simplest living systems (such as unicellular organisms) use metabolic pathways enclosed in cell membranes to maintain themselves while drawing materials from the environment. More complex living systems exhibit a much wider behavioral repertoire because groups of cells form organs with complex coordinated functions.
• Adaptivity is now not only achieved using chemical means but by changes in behavior, such as heavier breathing when oxygen content is lower or slower movement when it is very hot. Behavior is controlled using special-purpose neural networks.
• The most important innovation is however the preservation of information by coding the system in terms of genes. This requires the `discovery' that proteins can function as interpreters of a code [5]. The code itself, in the form of DNA, is now copied as opposed to the whole organism. Additional proofreading while copying assures that much more complex information can be preserved, not only for creating the next generation of an individual but also for regenerating constantly parts of a single individual.
• The genetic mechanism provides also a much more powerful way to generate more complexity. The code is mutated or combined via cross-over operations and then subjected to naturally occurring selection. A larger search space of possible life forms can thus be explored and it becomes easier to build further upon existing complex forms. Other ways are used to increase complexity as well, they include level formation and self-organisation. Based on these principles living systems have shown several transitions towards ever greater complexity. Recent overviews of the important transitions have been given by Maynard-Smith and Szathmary [12] and de Duve [3].

Intelligent systems can be defined as systems that have the same four properties (self-maintenance, adaptivity, information preservation, and increase in complexity) but use other means to achieve them. It is not yet completely obvious where the key lies, but two things are surely important:

• Neural networks, which initially were completely specific, have become general purpose structures which can store a large number of complex behavioral patterns, sustain processes for interpreting signals from the world and controlling at a fine grained level complex action patterns. Most importantly these networks and processes develop and adapt themselves continuously and very fast (compared to genetic evolution).
• At some point a symbolic capacity has developed: This is the ability to interpret the world in terms of concepts, to represent states of the world using these concepts, and to perform symbolic reasoning by manipulating these representations. This symbolic capacity also sustains symbolic learning.