First, we'll try to understand what “intelligent behavior” could mean
Then we'll take a look at the disciplines that deal with intelligence, cognitive science and artificial intelligence and try to understand what “models of intelligent behavior” are
What is intelligence?
What does “intelligent” mean? (from Latin: interleggere meaning “to understand”, lit. “to read between [the lines]”)
When is a system intelligent? (ontological questions)
How do we know when a system is intelligent? (epistemological question)
Question: intelligent = cognitive? (Is it possible to have intelligent systems that are not cognitive?)
Common sense notions include thinking, problem solving, learning and memory, emotions, consciousness, intuition and creativity, language capacity, etc.
One possible answer: whatever the IQ test measures (cp. to the Bell Curve by Herrnstein and Murrey, or the recent discussions on the G-factor)
Problem: limits intelligence to humans
Others:
“the ability to carry on abstract thinking”
“the ability to adapt oneself to the environment
“the capacity to acquire capacity”
“the capacity to learn or profit by experience”
“a biological mechanism by which the effects of a complexity of stimuli are brought together and given a somewhat unified effect in behavior”
Turing's approach: avoid definition of “intelligence” (Turing test—what does it measure?)
Also: emotional IQ, “street smart”, “(sporadic) good thinking”, etc.
What about animals? Are they intelligent? How can we measure their intelligence?
We assume there to be some sort of ordering relation...
Maybe there is no general answer to the question (maybe “intelligence” is a “cluster concept”...)
So:
first of all, avoid answer and focus on behavior, i.e., of behavior in particular circumstances
second, we are interested in how a given behavior can come about, how it can be implemented or how it can “emerge” from other behaviors
First, have to define what "cognitive science" means
Careful: there are probably as many different answers as researchers in the field
So I won't attempt to give you a clear-cut definition, but rather a few intuitions that will later crystallize as we study different models
Taken literally: science of cognition (from Latin: pp. of cognoscere meaning "to become acquainted with", "to know")
From The Mind´s New Science-A History
of the Cognitive Revolution? (Basic Books, New York, 1985) by Howard
Gardner: "Cognitive science is a contemporary, empirically based effort
to answer long-standing epistemological questions?"
Cognitive science as empirical approach to epistemology (that is, the theory dealing with knowledge, what it is and how it is possible)
What is “knowledge”?
implicit vs. explicit
positive vs. negative
declarative vs. procedural
innate vs. acquired
direct vs. indirect
Other definitions stress more the notion
of "information" and view cognitive systems as "information processing systems"--what
does "information" mean? (Shannon, Dretske, Barwise, etc.)
Other variant: interdisciplinary effort
to study and understand cognition--but what is cognition?
Other version: interdisciplinary effort
to study and understand cognitive systems--what are cognitive systems?
Yet another one: interdisciplinary effort
to study intelligent systems (again,
what are intelligent systems?)
Examples of cognitive systems: “only” (?) biological systems (not all animals)
What about “Deep Blue”?
Difficult issue, no clear unanimous answer available (the “model muddle”, see Barbara Webb's BBS article at http://www.bbsonline.org/Preprints/Webb/)
First, let's collect some basic properties of models:
models are abstract (they might have concrete instantiation, though)
models are simpler (than the target they are supposed to model)
models capture some relevant aspect of the target system
models are thought to elucidate crucial aspects of the target system
they show what it does
some models even show how the target system does (implementation!)
Different kinds of models;
mathematical models (e.g., system of differential equation)
physical models (e.g., a mass-spring system)
Both may or may not be computational
Other factors:
real-time vs. abstract or simulated time
real vs. virtual
Let's try: computational models are models such that they use computers in one way or another to model one or more aspects of a cognitive system
“to model” means to have the computer exhibit the same behavior under the same conditions as the cognitive system (“same” is tricky!)
For the books: computational models are models that are
computable (in the sense of the theory of computation)
implementable on standard computers
and hopefully computationally feasible
Distinguish: “computer simulation” from “computer model”
the former “simulates” something else which might or might not be a computer model, whereas the latter clearly is a model qua computer
Very interesting issue: what is the difference between “simulation”, “emulation”, and “implementation”
another term in this context is the philosophical notion “supervenience”
What are we going to implement in the course?
Sensory and motor routines
Aspects of cognitive systems
Behaviors, etc.
Possibly psychological, biological models (in which case we need to build on results from psychology, biology, neuroscience, ethology, etc.)
Engineering approach to intelligent systems:
try to mimic human/animal behavior (regardless of whether it is biologically plausible)
Cognitive science approach:
do it in a biologically plausible way
=> Criterion: not only the input-output behavior must be right, but there has to be an additional structural relation between the target system and the model:
Some sort of “internal, structural correspondence”
What are the properties of such a correspondence?
At what level does this correspondence have to be established?
How “far down” does it have to/need to go?
For this course: try to get things to work (inspirations from the behavioral sciences are sufficient, don't need biological plausibility, although it is a desired feature)
First note that AI and CogSci overlap
In particular, cognitive modelling and AI overlap to a large extent, but they are not identical
there is an engineering side of AI as well!
Common interests:
To model human or animal behavior
To model human thought
To implement rational thought
To implement rational action
Models in AI and cognitive science usually are the result of sequence of
behavioral experiments
the development of computational model
A caricature of the various the interplay between AI, psychology and philosophy within cognitive science (taken from M. Gasser—note: the speakers below are completely fictitious are not meant to resemble any person, living or dead):
Philosophers and psychologists on AI
Philosopher: "You people are playing with notions like concept and meaning as though you invented them. We've been dealing with this stuff for more than 2000 years; you might benefit by reading something so you don't re-invent the wheel. And frankly you seem a bit too cocky to me. Programs that 'understand' natural language? Give me a break."
Psychologist: "I'll be blunt; I don't understand what you people are 'modeling'. Where's your data? You'd benefit by doing some experiments on real people so you'd have some way of knowing whether your 'models' are on the right track. As for as I can tell, all this talk of 'cognitive modeling' is just a way to dress up engineering, which is fine, but please don't pretend it's cognitive science."
AI people and psychologists on the philosophy of mind
AI worker: "To be honest, I don't know how to make sense out of what you people are saying. I mean, half the time I don't have a clue how I'd take your ideas and turn them into an algorithm. What good is all this if we can't use it to write programs so we can test whether it makes any sense? There's a place for the humanities, but not in cognitive science, please."
Psychologist: "All this sounds very nice (when I can understand it), but isn't it all just a bunch of armchair speculation? You'd be a lot better off basing it on some real empirical research, that is, if what you're claiming is testable at all, which I doubt. And another thing; can't you bring some transparencies when you give a talk? I'm used to looking at text and graphics."
AI people and philosophers on psychology
AI worker: "It's all well and good to spend your time in the lab gathering data. I have lots of respect for real hard numbers. But most of the time your results are things I already pretty much knew were true. We really don't need your help to build our models. There's plenty about intelligence we already know that's crying out to be modeled. And the quantitative models can wait till we solve the more interesting qualitative stuff. And where are your theories anyway? You spend 95% of your time gathering data and doing statistics on it, but what about an explanation for all this?"
Philosopher: "I know you people
would like the study of cognition to be a real empirical science like
physics, but you seem to spend all your time focusing on trivia, like
how many milliseconds it takes people to push a button in response to
some kind of image on a computer screen. What about the big questions?
And how can you do your experiments in an intellectual vacuum? Do you
really begin to understand all the hidden assumptions behind what you're
doing?"
IMPORTANT: cognitive science attempts to overcome interdisciplinary difficulties such as
different aims
different vocabulary
different background knowledge
different modelling techniques
The role of learning
If you want an intelligent machine, program in the intelligence.
If you want an intelligent machine, make it a good learner and send it out into the world.
The role of the hardware
Intelligence is a software problem. An intelligent program can be run on a brain or a computer.
The hardware is relevant: we should look for intelligence that is based on the properties of nervous systems. [neural networks, connectionism]
The role of domain-independent methods
Neat AI (mostly overlapping with Haugeland's GOFAI-"good old-fashioned AI")
Theories should be elegant and parsimonious.
We should understand precisely what our theories can do and how they behave.
Most (or all) of intelligence is governed by general principles.
Scruffy AI (mostly overlapping with Haugeland's NFAI-"new-fangled AI")
The mind is a kludge. To make things efficient, inelegant short-cuts are often appropriate.
It may be impossible to come to a precise understanding of our theories.
There are only a few general principles that apply across domains. Intelligence comes from domain knowledge.
Two main paradigms: symbolic and non-symbolic
amounts to "classical computational" and "dynamical"
CAREFUL: the above "computational" does not literally mean computational, but refers to the GOFAI style of modelling
Three basic modeling frameworks:
symbolic
connectionist (sometime also called subsymbolic)
dynamical
“Behavior-based robotics” is a cross-cut in some sense (will talk about the tenets of BBR later)
Other distinctions to follow as we proceed (e.g., embodied, situated, etc.)
No general consensus about the exact day of birth
2 major (founding) events:
the Hixon Symposium on cerebral mechanism in behavior (1948)
interdisciplinary meeting of researchers interested in cybernetics and the theory of information (e.g., John von Neumann, Warren McCulloch, and Norbert Wiener)
combine ideas from mathematics, the new field of "computer science", cybernetics, and neurobiology
e.g., discussions of the "neuronal model" and its informationtheoretic properties
the symposium on information at MIT (1956)
among the participants Noam Chomsky, Claude Shannon und George Miller
Chomsky's “colorless green ideas sleep furiously” (as opposed to “ideas furiously green colorless sleep”)
syntactic, grammatical, rule-based structure
Shannon's “syntactic” notion of information (information content of a signal is the inverse of its probability of its expectation)
Miller: the magic number 7 in memory tasks; writes about the second day of the meeting that
“cognitive science burst from the womb of cybernetics and became a recognizable, interdisciplinary adventure in its own right” (Miller, A Very Personal History, Cambridge, MA: MIT Center for Cognitive Science, 1979, S. 9).
Furthermore, Miller wrote that he left the symposium with the conviction that “human experimental psychology, theoretical linguistics, and the computer simulation of cognitive processes were all pieces from a larger whole, and that the future would see a progressive elaboration and coordination of their shared concerns” (Miller, 1979, S. 9).
Also: Marvin Minsky and John McCarthy's workshop in the summer of 1956
Herbert Simon and Allen Newell present Logic Theorist, the first prototype of an artificial intelligent system
the term “artificial intelligence” is coined by McCarthy
Marvin Minsky states in the proposal to the 1956 Dartmouth conference:
“an intelligent machine would tend to build up within itself an abstract model of the environment in which it is placed. If it were given a problem it could first explore solutions within the internal abstract model of the environment and then attempt external experiments.” (McCarthy, 1956)
Claim: cognitive science would not have been possible without the rise of the digital computer!
From today's perspective clear: humans are able to process (complex) information (i.e., receive, store, retrieve, transform, etc. information)
Note possible to say in the 40ies and beginning 50ies in psychology
Behaviorism dominated the field in the US, cognitive psychology was regarded “unscientific” (since it had used introspection as a method to obtain knowledge about the mind)
Computer helped: computers also process
information (just like other biological systems), but furthermore we know
how they do it! (e.g., computational states = inner states of the
system; almost abhorred by behaviorists)
To compute a function on a computer (i.e., to let it exhibit a certain behavior), it suffices to specify the program description of the function (i.e., the algorithm)
=> specify various computational states
All these ideas (computer as information processing system, syntactic specification of programs that give rise to a causal chain of actions inside of the physical system “computer”) suggest a daring analogy:
maybe cognition (understood as the sum of all information processing cognitive processes) are nothing but a complex computational process?
Put differently: maybe cognition is computation?
“The mind is to the brain as the program is to the hardware?” (z.B. Searle, Johson-Laird, Dennett, Phylyshyn, u.a.)
Contributed essentially to the rehabilitation of cognitive psychology as a scientific discipline and as the “midwife” assisting the birth of cognitive science
(Still) main paradigm in cogsci
The program based on it is sometimes called “computationalismus?" or "the computational claim on mind” (“cognition is computation and can be understood as such”)
Computationalists view cognitive processes as computational manipulation of representations (i.e., syntactic transformations of discrete entities, so-called “symbol tokens, within the cognitive system, which might stand in for something within or outside of the system
In other words: these symboltokens are representational!
Cp. to computer model:
in digital computers there are discrete units (e.g. bits), which are manipulated and transformed during a computational process which is syntactically specified by a program
In philosophy of mind: “Turingmaschine”-functionalism
The computationalist claim: cognitive processes can be described syntactically without paying attention to their semantics (vide Haugeland: “semantics will take care of itself if only the syntax is right”)
Fodor (1983) is one of the strongest proponents of the modularity hypothesis
3 level model:
transducer
input/output system (modular)
higher cognitive functions (non-modular)
(note that some systems can indeed be clearly distinguished from others)
Modules are:
functionally specific (their computational organisation depends on the respective stimulus)
cognitive impenetrable (e.g., one has no conscious control over how one perceives things)
fast (since they work autonomous)
informationally closed (do not need any information of other systems to do their job)
flat output (no complex representations)
Two major frameworks:
symbolic
non-symbolic (dynamic, etc.)
Distinguish: symbolic, connectionist, and dynamic (each may or may not be computational)
Do not equate symbolic with computational!!!
(although most computational models are symbolic models)
Symbolic models
Physical Symbol Systems (Newell, Pylyshyn, Fodor; summarized by Harnad)
A set of arbitrary physical tokens (scratches on paper, holes on a tape, events in a digital computer, etc.) that are
manipulated on the basis of explicit rules that are
likewise physical tokens and strings of tokens. The rule-governed symbol-token manipulation is based
purely on the shape of the symbol tokens (not their "meaning"), i.e., it is purely syntactic, and consists of
rulefully combining and recombining symbol tokens. There are
primitive atomic symbol tokens and
composite symbol-token strings. The entire cognitive system and all its parts--the atomic tokens, the composite tokens, the syntactic manipulations (both actual and possible) and the rules--are all
semantically interpretable: The syntax can be systematically assigned a meaning (e.g., as standing for objects, as describing states of affairs).
Processes happen sequentially.
There is a central controller which coordinates the activities of the modules of the cognitive system and selects among candidate processes at each point in time.
The cognitive system interacts with the world through interfaces to perception and action, which operate very differently from the internal (cognitive) system.
Knowledge is usually programmed into the cognitive system by someone who has a theory of how knowledge is organized. Learning is also possible, but it is not central to most models.
Time is often mapped onto space; that
is, the cognitive system has simultaneous access to all of a pattern of
some length (word, sentence, etc.). Inputs may also be presented sequentially,
but the problem of temporal short-term memory is side-stepped because the
inputs are preprocessed
Non-symbolic (maybe "sub-symbolic", connectionist, dynamic, etc.) models
Control is distributed. There is just the illusion of someone being in charge because the behavior seems purposeful, and it seems to be possible to write a centralized program to make it happen.
The basic processes involve very simple interactions among primitive elements arranged in a network. Usually the interaction amounts to the spread of activation.
Many of the processes happen in parallel.
The cognitive system may interact with the world through perception and action components which are similar to the internal (cognitive) parts of the creature. In some models, the environment and the creature itself constitute one large dynamical system.
Except in localized connectionist models, knowledge is distributed, usually in the form of patterns of connectivity among the primitive elements. The knowledge in such systems is implicit; it often cannot be simply read off.
Knowledge gets into the cognitive system (except in most localized models) through learning as the system discovers the statistical properties of the world around it or through evolution as generations of creatures are forced to survive in the world.
The problem of temporal short-term memory is often addressed, though the continuous interaction of components of the cognitive system with each other and the world may not be. In the most conservative approaches, input patterns are fed to the cognitive system in the form of a sequence of discrete events. In the most radical approaches, the cognitive system exists in the world in continuous time.
Early to Mid-Eighties (upcoming connectionism, Rumelhart & McClelland's PDP, 1988)
Also, AI did not reach expectations
First criticism of the computational model
Note philosophers had been criticizing it for quite some time
the Gödel-Lucas argument
Searle's "Chinese Room" thought experiment
Common tenor of critics: the symbolic, top-down model might be insufficient
Maybe implementation details matter!
Maybe natural intelligent systems are not computational!
So, first look at the neuronal level
to figure out what these neurons do, then start contemplating about what
"architecture" they implement
Connectionists: researchers who study the properties of neural networks (with respect to cognitive functions)
Main convictions:
symbols cannot be assumed,
they need to be viewed as "emergent" upon the properties of neural/connectionist networks
use biologically plausible models (such as neural networks)
The rise of the neurosciences (as indicated by the "neuro"-prefix):
Neurolinguistics
Neuroinformatics
Cognitive Neuropsychology/physiology
Beginning of the 90ies: even more radical approaches against the classical model
Dynamicists (=adherents of the theory of dynamical systems)
criticize not only the classical level of description and/or its various methods of implementation
also doubt that representation is the central notion
Good source: It´s about Time“, byRobert Port and Tim van Gelder, MIT Press, 1995.
Similar development in AI: embodied AI
Advanced mainly by roboticists (Rodney Brooks, Hans Moravec, et al.)
Main credo: cannot study mind without body
Claim: many cognitive processes do not have to be implemented directly, but will result as a byproduct from the interaction of the agent with its environment!
Important: physical (not simulated) agents as physics matters (i.e., the physical properties of the body)
=> Overlap between dynamicists and embodied AI people
Binding Problem
Frame Problem
Symbol Grounding Problem
Innate Knowledge
Innateness and Emergence
Intentionality
Levels and Cognitive Architectures
Modularity
Representations
Representation and Computation
Rules
Stages of Processing
Pfeiffer's notion for cognitive science that deals with "complete agents"
Two kinds of agents
real (in the sense of "physical")
simulated
What is the difference?
Should real agents be preferred?
Basic notions:
Self-sufficiency
multiple tasks and behaviors
trade-offs and deficits
circadian cycles
behavior control
Autonomy
degrees of autonomy
dependence on the environment vs. dependence on other agents
self-sufficiency increases autonomy
Situatedness
Definition: "An agent is situated if it acquires information about its environment only through its sensors in interaction with the environment"
agents can acquire their own history
more autonomous
evolutionary techniques
Embodiment
agents have a body, they are physical agents (merely simulated agents can never be embodied!)
embodied agents are forced to interact with their environment
agents are subject to physical forces, etc.
possibility to utilize physics for control (e.g., gaits, etc.)
Adaptivity
the ability to adjust oneself to the environment
adaptivity and intelligence are directly related
different kinds:
evolutionary adaptation
physiological adaptation
sensory adaptation
adaptation by learning
Ecological Niche
Definition: "The range of each environmental variable (such as temperature, humidity, food items, etc.) within which a species can exist and reproduce"
no universal animal! (some contrast this to "universal computation")
need to characterize the niche with respect to a particular agent:
only properties that are behaviorally relevant matter
static vs. dynamic environment
deterministic vs. non-deterministic
the sensors and effectors (i.e., how the system can interact with the environment)
the objects in the environment and their properties