Chapter 1: The Study Of Intelligence

Intelligence has always been a controversial, emotional topic. This book is about intelligence; but this term is very difficult to define. “There is very little agreement on what does and does not constitute intelligence”. Discussion of what intelligence is usually evolves into a discussion of ‘interesting behaviours’. “An exact characterization of intelligence is not all that important to understanding it.” NB: is this a logical or plausible position to take?

1.1 Characterizing Intelligence

Intelligence is hard to define, and as a result it is not surprising that there are many different definitions of it. Pfeifer and Scheier prefer to explore some commonsense properties of intelligence to get a better handle on it (NB: see issue raised above – is this a consequence of it?). From this common sense perspective,

Intelligence is a graduated property that varies in degrees
Intelligence often seems to involve thinking, problem solving or abstract thinking
Intelligence is reflected in the ability to learn and to remember, but not in rote learning
A hallmark of intelligence is the ability to communicate in a natural language
The highest form of intelligence might be creativity, but creativity might also be distinct from intelligence
Consciousness (an elusive subject in its own regard) may be an essential component of intelligence
Intelligence might have an emotional component
Sophisticated behaviours for surviving in a complex world reflect an underlying intelligence
Perception is not a component of intelligence from a commonsense perspective, but is still a sophisticated ability

An alternative approach to understanding or defining intelligence comes from the testing and measurement tradition; can intelligence be reduced to a single number? Pfeiffer and Scheier do not believe so. Intelligence testing leads naturally to the nature vs. nurture debate – to what extent is knowledge or intelligence innate? There are tremendous social policy issues that hinge on this discussion. Also, the possibility of machine intelligence emerges from the consideration of the Turing test.

The Turing test is itself very controversial. In many respects, it is very restricted and not very powerful – hence Searle’s Chinese Room argument. But isn’t this test the basis for all our assessments of intelligence – eg., our assessment of whether Searle is intelligent? (NB: Penrose position is related to this).

With all of this diversity, is there a common denominator for defining intelligence? “There seems to be one underlying common theme that involves ‘coming up with something new’ … Surviving in the wild means coping with novel situations which in turn implies behaving in new ways”. This leads to the diversity-compliance trade-off: “What are the mechanisms enabling organisms to adapt to, cope with, environmental changes? As we noted, adaptation always involves two components: complying with existing rules and generating new behaviour; only if both components are present do we speak of adaptivity. It then makes sense to tie intelligence to adaptive behaviour”. (my italics) The key implicit ingredient to all of this is embodiment. NB: we could pursue this in a pro vs. con argument.

1.2 Studying Intelligence: The Synthetic Approach

How to proceed? First, distinguish analytic from synthetic approaches. “By contrast, the synthetic approach works by creating an artificial system that reproduces certain aspects of a natural system. This is another important function of models. Rather than focusing on producing the correct experimental results, that is, the correct output, we can try to reproduce the internal mechanisms that have led to the particular results.” NB: key idea here is that data fitting is not the primary goal of simulation!!

Synthetic psychology is at the core of embodied cognitive science. Traditional AI biews intelligence as being closely akin to what computers do – the information processing metaphor. This view has had many important successes – but not in building robots to work in the real world!

Solution – abandon the traditional view of intelligence. Instead, view intelligence as emerging when loosely coupled and parallel systems interact with each other and with the real world. The synthetic approach will involve creating simulations, but will also involve the building of systems – autonomous agents – that exist in the real world. They will behave without the intervention of human control. Whether simulation or embodied agents are studied will depend on the situation, but both are likely to be required.

Some examples: Webb’s cricket work, other robot examples like Cog.

“A second application of autonomous agents in cognitive science is to explore principles of intelligence. This approach draws inspiration form nature, but offers us more freedom than the modeling approach. Experiment can be conducted using any type of sensor, even sensors that do not exist in nature”. Key idea here is notion of emergence. NB: -- what do you get for free?

With this synthetic approach, there is a big window of opportunity for exploring applications. Plus, the key issue of design begins to reign: “How would we design a system that behaves in a particular way that we find interesting?”