rmpfp@minyos.xx.rmit.oz (Frank Papa) (07/18/90)
Hi neural-netters, this is my first posting. I want to see what
others think of the following idea on organising AI/neural-network
research into a seemingly viable model.
Here goes:
"The presentation of a model for an intelligent process."
by Frank Papa.
The author intends to present a theoretical model of an intelligent
cognitive process to encourage discussion and better understanding in
this field.
It is also anticipated that this model will encourage researchers of
Artificial Intelligence to co-ordinate their specialised areas towards
a common understanding of:
- An intelligent processes' requirements,
- Developing the necessary structures which allow the process to operate,
- Interconnecting these structures,
- Driving the entire process.
The use of neural network theory is implied throughout the model,
including interconnectivity and localised memory.
The following is a build up to the simplified AI model
presented in figure 11. Diagram arrows represent a majority direction
of information flow, and include a smaller amount of information
flowing in the reverse direction.
"The conscious level."
A mobile intelligent process moves in a co-ordinated manner.
It uses trajectory planners to calculate motion paths.
These planners utilise hypothetical space modelling to command
actuators. Figure 1 shows a Hypothetical World Model supplying
commands to the Actuators. A common application of trajectory generation
is in the control of a robot arm.
Hypothetical ------> Actuators
World Model.
Figure 1
The hypothetical world model process is used for planning
appropriate activity (motion, speech), and for conceptualising.
Many separate models may exist within this process, allowing comparison
between possible actions, and allowing ideas to be followed. The
actuator functions do not directly modify the model, they are simply
responding to commands. Sensors indirectly provide actuator feedback
to the model.
Hypothetical world models are developed from planning algorithms,
with information taken from the Actual World Model. Abstract concepts
and unobservable events are extrapolated from memory. Planning is
subject to motivated goals, and to limitations set by laws. These
limitations include physical laws, grammatical laws and behavioural
laws. Figure 2 shows the Hypothetical World Model receiving information
from the Actual World Model.
Actual World Hypothetical
Model ------> World Model.
^
|____Planner
Figure 2
An Actual World Model is important as it feeds information to
the planning process and to the Hypothetical World Model.
World modelling processes use language to manipulate information
concerning location; geometric, mathematical, and physical properties;
associations, and interaction laws. This information is made up of
cognitive processes which manipulate, associate, or construct objects
and conditions within a model. These processes include audible, visual,
tactile, mathematical, logical, and algorithmic languages. Figure 3
shows the Actual World Model receiving Information and commands from via
Language.
Language ------> Actual World
Model
Figure 3
The attributes of language represent information which may alter
the world model. For example, being told that "The box fell off
the table" modifies the world model information pertaining to the box.
Information attributed to the box is modified to include that the
box is on the floor. This information follows from the physical laws
and constraints that the floor will stop the box's fall. Appended to the
box's history is a falling trajectory. The box is attributed with
the possibility of damage due to its collision with the floor.
The level of sound heard versus the known strength of the box's material
and contents qualifies the possibility of damage having occurred.
The information attributed to language is processed information
received from sensors. The fallen box example uses audio sensing, and
relies on memorised visual language to recreate the episode. Figure 4
shows Sensors supplying Language processes with raw information.
Sensors ------> Language
Figure 4
The five processes presented so far form the model's "conscious"
level. This is shown in figure 5.
Actual Hypothetical
Sensors---> Language ---> World ---> World ---> Actuators
Model Model
^
|____Planner
Figure 5
"The unconscious level."
The process described thus far would not pass the Turin test.
It can only respond to external stimuly in an unintelligent manner.
Planning is a goal oriented activity. These goals are provided
by external requests acting upon internal motivations. Figure 6 shows
the planner receiving motivational requests.
Motivation ------> Planner
Figure 6
Motivation comes in many forms to an intelligent process.
Internal motivations arise from a need to maintain the process. These
include requests for energy (hunger - eat food), survival (pain - fight,
flight), and procreation.
External requests encourage, or inhibit these internal
behavioural goals. Consider the number of times an employee will endure
hunger in order to complete a piece of urgent work. Figure 7 shows the
motivation process receiveing maintenance requests.
Maintenance ------> Motivation
Figure 7
The two levels presented so far form the conscious and unconscious
levels of the model. These two levels are shown in figure 8.
Actual Hypothetical
Sensors---> Language ---> World ---> World ---> Actuators
Model Model
^
|
Maintenance ---> Motivation ---> Planner
Figure 8
"The driving level."
The process described thus far would correspond to a intelligent
process only capable of responding to external demands and basic internal
needs such as energy. When neither stimuli source is present, the process
would sit idle.
The Earth presents an ever-changing environment. Most of Earth's
inhabiting creatures spend nearly all their adult lives responding to
external stimuli and internal maintenance requests. The complexity of
human social life results in only an acute awareness of each individual's
driving level.
The driving level provides the intelligent process described so
far with a perpetual momentum, so that it never lays idle. It also
creates the motivation necessary for learning, which is needed by an
intelligent process to cope with its environment.
The driving level feeds the motivation process. Unlike the
maintenance requests, driving level requests (called requestors) begin
their life as random requests which are either strengthened through
success, or dismissed through failure. The purpose of their requests
is to diversify the range of goals set in order to satisfy the demands
of the paradox loop. A diagram showing requestors feeding the planner
is shown in figure 9.
Planner
^
|
Requestors
Figure 9
The end result of behaviour originating from the driving level
appears to an outside observer as if the intelligent process is behaving
in a creative manner. Actions and concepts arise from the intelligent
process which seem both to an observer and to the process itself, as
original. These externally observed phenomena emanate from the world
models through to the actuators. As a result they are limited to the
processes represented by the models.
For example, an intelligent process may decide to follow and
enquire about a moving object. This curious behaviour might have no
relevance to external stimuli (except that the object has been sensed
and its geometric description is stored in the world modeller), nor
internal maintenance requests. The request for information concerning
the object will have emanated from the requestors in order to satisfy
the paradox loop.
An important point to make here is that the intelligent process
includes a description of itself in its models. Its description of
internal requestors is limited to language descriptions of "curiosity",
and "creativity".
The paradox loop perpetuates the activity of requestors by
requesting solutions to an unsolvable puzzle (eg: The next sentence is
true. The preceding sentence is false). Fiogure 10 shows the paradox
loop feeding the requestors process.
Requestors
^
|
Paradox Loop
Figure 10
A complete model of the intelligent process is presented in
figure 11.
Actual Hypothetical
Sensors---> Language ---> World ---> World ---> Actuators
Model Model
^
|
Maintenance ---> Motivation ---> Planner
^
|
Requestors
^
|
Paradox Loop
Figure 11
This complete model forms an inverted pyramid with exponentially
increasing complexity. Each block represents a specific task with
protocols to each other task.
____________
\ | | | | |/
\ | | | |/
\ | | |/
\ | |/
\/
Figure 12 Inverted Pyramid Structure.
Arranged in these groups are neural cells. Is this arrangement
by chance or is it learnt? A certain amount of interconnectivity is
preprogrammed. This occurs only at the very fundamental levels. Most
interconnectivity, which forms the mechanisms for processes abilities,
is a result of the cognitive manipulation of incoming data and through
trial and error. This is identical to the formation of requestors, whose
interconnectivity strengthens or dies depending on location and use.
Conclusion.
It is hoped that this model will direct AI researchers,
especially those involved with neural networks, to organise their
contributions so that it fits in with this seemingly workable
arrangement. The most obvious task is to find an application simple
enough to test this model.
I would appreciate any comments and suggestions.
Please direct mail to Frank Papa, VUT-RIMT
(Victorian University of Technology -
Royal Melbourne Institute of Technology)
GPO Box 2476V, Melbourne, Australia, 3001.
phone (03) 660-2168.
on email: rmpfp@minyos.xx.rmit.oz
Thanks.