=======================================================================


                  Abstract for I3 Spring Days Workshop
                                   on
         BEHAVIOR PLANNING FOR LIFE-LIKE CHARACTERS AND AVATARS

                       Architectural Requirements
                         for Human-like Agents
                   (What sorts of machines can love?)

                              Aaron Sloman

                       School of Computer Science,
                      The University of Birmingham
                           Birmingham B15 2TT
                                   UK
                      EMAIL A.Sloman@cs.bham.ac.uk
                     http://www.cs.bham.ac.uk/~axs/

Abstract:

If instead of thinking only about normal adult humans we consider
also infants, people with brain damage or disease, and also other
animals including insects, bacteria, birds, bonobos, etc., we
find evidence for myriad information processing architectures each
supporting and explaining a specific combination of mental capabilities.
Yet more possible architectures, each supporting a collection of
possible states and processes can be found in robots, software systems
and machines of the future!

Thus concepts describing mental states and processes in one animal or
machine may be inappropriate when describing another. Likewise, concepts
relevant to normal adult humans may be inappropriate for new-born
infants or victims of alzheimer's disease.

It is arguable that much of what poets and novelists say about us, and
what we say about our friends and ourselves when gossipping or discussing
our interests, loves, hopes, fears and ambitions, implicitly presupposes
that humans are essentially information processing systems. E.g. when
poets distinguish fickle liking which is easily diminished by new
information and deep love which is not, they implicitly presuppose that
new information can have effects on powerful information-based control
states.

By considering possible descriptive and explanatory concepts generated
by a *virtual machine based information processing architecture* we
obtain a broader and deeper explanatory theory than is normally found in
philosophy, psychology or social science.

A proper understanding requires comparative analysis in design space. We
understand a particular architecture better if we know what differences
would arise out of various sorts of design changes: which capabilities
would be lost and which would be added. This involves going beyond the
majority of AI projects or psychological investigations in considering
both designs for *complete* agents and also doing *comparative* analysis
of different sorts of designs.

Discovering the architecture of a complex system we have not designed
ourselves is very difficult. No amount of observation of the behaviour
of any animal or machine can determine the underlying architecture,
since in principle any lifelong set of behaviours can be produced by
infinitely many different information processing architectures.

Decompiling information gleaned from either invasive or non-invasive
observation of internal physical structures is just as hard, e.g. if we
don't even know at what physical level most of the architecture is
implemented. Do neurones or molecules do most of the information
processing?

We can best constrain our theories by combining a number of
considerations, such as: (1) trade-offs that can influence evolutionary
developments, (2) what is known about our evolutionary history, (3) what
is known about human and animal brains and the effects of brain damage,
(4) what we have learnt in AI about the scope and limitations of various
information processing architectures, mechanisms and representations,
(5) introspective evidence, such as my knowledge that I considered and
evaluated alternative ways of travelling to this conference.

But our theories are will be condemned to being *conjectures* for years
to come. At least we can show that some conjectures are better than
others, if we take a broad enough view of what needs to be explained.

I'll offer a conjectural theory ((a)-(f) below) based on those
constraints:

(a) Evolution, like engineers, found that (partly) modular designs are
essential for defeating combinatorics in the search for solutions to
complex problems (with only 4,000,000,000 years and one biosphere).

(b) Human information processing makes use of (at least) three different
concurrently active architectural layers which evolved at different
times, which we share with other animals to varying degrees, and which,
along with various additional supporting modules, account for different
cognitive and affective states, as well as different possible effects of
brain damage, and other abnormalities.

(c) Reactive, deliberative and reflective layers support different
classes of emotions found in humans:
    (i) the reactive layer accounts for primary emotions (e.g. being
        startled, frozen with terror, sexually aroused);
   (ii) the deliberative layer supports emotions like apprehension and
        relief which require "what if" reasoning abilities;
  (iii) a meta-management (reflective) layer supports not only control
        of thought and attention but also loss of such control, as found in
        typically human tertiary emotions such as infatuation, humiliation,
        thrilled anticipation of a future event. (It is also crucial to
        absorption of a culture and various kinds of mathematical,
        philosophical and scientific thinking.)
All the layers are subject to interference from the others and from one
or more fast but stupid partly trainable "global alarm" mechanisms.

(d) Perceptual and motor systems are also layered: the different layers
evolved at different times, act concurrently, and have different
relationships to the "central" layers.

(e) Analysing ways in which components of such an architecture might
bootstrap themselves, develop, reorganise themselves, acquire and store
information, will provide far richer theories of learning and
development than ever before.

(f) A multi-layered architecture of the sort proposed could give
robots human-like mental states and processes, including qualia, leading
some of them to re-discover philosophical confusions about
consciousness. Software agents could have similar capabilities.

Many doubt this because they see the limitations of existing
computer-based machines and software systems and cannot imagine any ways
of overcoming these limitations. They do not realise that we are still
in the early stages of learning how to design information processing
systems. (Claiming that computers will be ever more powerful is not
enough to allay these doubts: we also need deep analysis of the concepts
used to express the doubts.)

Existing AI systems do not yet have whatever it takes to enjoy or
dislike doing something. They do not really *want* to do something
or *care* about whether it succeeds or fails, even though they may
be programmed to give the superficial appearance of wanting and caring.
Simulated desires and emotions represented by values for global
variables (e.g. degree of "fear") or simple entries in databases linked
to condition-action rules fail to address the way emotions emerge from
interactions within an architecture, and fail to distinguish emotions
requiring different sorts of architectures.

Current attempts to replicate other animal abilities are also limited:
for example, visual and motor capabilities of current artificial systems
are nowhere near those of a squirrel, monkey or nest-building bird. To
understand animal comprehension of space and motion we may need to
understand the differences between precocial species born or hatched
with considerable independence (chickens, deer) and altricial species
which start utterly helpless (eagles, cats, apes). Perhaps the
bootstrapping of visuo-motor control architectures in the latter yields
a far deeper grasp of space and motion than evolution could have
pre-programmed via DNA. Artificial agents which do not share our deep
grasp of spatial structure and motion will be limited in their ability
to communicate with us.

Depending on time available and the interests of people at the workshop,
these and other related themes will be presented and their implications
explored.

Acknowledgement: many colleagues and students have helped with the
development of these ideas, most recently Brian Logan.

NOTE:
This work is presented at length in papers in the Birmingham University
Cognition and Affect FTP archive. Recent papers are listed in
    ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect/0-INDEX.html
The project is summarised in
    http://www.cs.bham.ac.uk/~axs/cogaff.html

Our software tools including the SIM_AGENT toolkit (Pop-11 based
code and documentation) can be found via this directory:
    ftp://ftp.cs.bham.ac.uk/pub/dist/poplog/