MRI 3 - The Foundations of a Foundation

MRI 3.1 Organising the Foundations of a Foundation
---------------------------------------------------

In this document we need to do some consolidation by laying out a foundation
of principles and ideas from what research we have gathered, and the ideas
which we have come up with so far.

This foundation will have to consist at least, of the following components:

1. A set of principles that provide the vision of our kind of intelligent
system

2. The basic principles and ideas of our approach, and the rationale behind
them

3. What questions remain to be answered, and what needs to be done to answer
them before we can at least start building something that fulfils our vision
partially


We will try to fill up those categories shown above, taking exerpts from MRI
1 and 2.


MRI 3.2 Our vision of an intelligent system
--------------------------------------------

"Intelligence" includes (ie. not exhaustively):
1. The ability to execute difficult tasks efficiently
2. The ability to adapt to increase the efficiency in tackling these tasks

The intelligent system must have the following qualities (...some apart from,
others in assistance in...) fulfilling the above two criteria:

1. Decentralised operation - everything is everywhere and nowhere...and in no
particular place.

2. The architecture breeds its own intelligence. Hence the complexity of its
behaviour comes from the operation of the system.
Our design is concerned with maximising the potential of the system to
generate complex behaviour. (Thus for a robotic organism living in a highly
dynamic environment requiring complex behavioural mechanisms, the system can
provide that)

3. Precision is not always necessary. The system should be tolerant of
inprecision. But we should take a step further ahead of nature, and demand
that the system should be capable of producing precision when required.

4. The system should be adaptive. It should learn and optimise itself.
It should posses the skill of learning, which we said in MRI 1.




MRI 3.3 Basic Principles and Ideas of our Approach and their Rationale
-----------------------------------------------------------------------

1. The Intelligent System is composed of a network of interconnected units
whose behaviour follows certain rules.
These are the key features of the biological neural network, which we believe
are directly responsible for the desired characteristics which we specified
in the previous section. Hence we have reason to investigate these features
more thoroughly and incorporate them in our nascent intelligent system.

The underlying simplicity of such an architecture is also attractive to us,
yet there is empirical evidence that this form has potential to generate
complex behaviour.

As the system derives its functionality in a decentralised manner, the
resulting robustness is also very desirable.

2. In order to exhibit the two important qualities that characterise
intelligence (as stated in the previous section), we need to create a source
of power or motivation for the system to do so. We call these DRIVEs.

We propose two drives, the EMOTION drive (E drive) and the SURVIVAL drive (S
drive).

The S drive is based on the biological organism's need to survive. To
survive, certain physiological parameters must be satisfied. These parameters
are fed into the system, which seeks to act via its outputs to ensure the
parameters stay within certain prescribed limits. In biological organisms,
these limits probably evolved. In our robotic organisms, we can choose to set
these limits ourselves in our designs, and perhaps allow those limits to
evolve through successive generations of robots (refer to the document on
robocology).


The E drive is based on emotions in biological organisms, most obviously in
humans. Emotions come in pairs, positive and negative ones. The system seeks
to minimise the negative emotions and maximise the positive ones.

The E drive may help complement the S drive in ensuring survival.
In addition, its purpose may also help increase the variety and complexity of
behaviour the system can exhibit.
Perhaps emotions evolved to enhance the survivability of biological
organisms. Perhaps some emotions like anger and fear were evolved to enhance,
but the nature of emotions in a network architecture meant emotions were dual
and had to come in pairs. Hence came the positive emotions. And perhaps at
first this served as a drive to keep the emotional state positive, since
negative emotions implied a threat to survival and actions had to be taken to
avoid this. Then the dependance on positive emotions evolved from this.

We propose that, similar to the S drive, what condit\ions evoke positive
emotions in the E drive mechanism can be set by the designer, and then
allowed to evolve over generations of robots.


3. In the conception of robocology, we seek to create an environment that a
robotic organism would survive in. The environment creates the conditions and
the challenge, while the DRIVES created the driving force for the robot to
behave in a way that would meet the challenges of the environment.

The driving force alone is not enough. It must drive some 'machinery'. These
are certain specifics of behaviour and skills to design. We draw inspiration
from subsumption, to break down certain broad tasks (which we must first
define) such as motor control and visual perception into smaller, more
manageable bits and layered into hierarchies of control.


4. These 'modules' may perform highly specific functions, such as those
performed by the lowest level pattern generators in the brain's motor control
hierarchy. They merely 'execute'.

Other higher level modules, or groups of them or the whole layer of modules
interacting with one another might perform the additional tasks of 'creation
and conception', and also 'learning and self-enhancement'. These fulfil our
two criteria of intelligence stated in the previous section.

while conception could have given rise to the patterns of behaviour exhibited
by the lowest level 'modules', this would have taken too much time to evolve,
since their effect is least profound, and to hit upon the right combination
of behaviour among a large number of modules to meet the high level of
objectives provided by the S and E drives takes too long to happen.


5. We can generalise every task to be a task of
1) conception of objects
2) destroying or creating connections between sets of objects

In other words we have incorporated the connectionist model of cognition, and
extended it to all parts of the intelligent system from low level behaviour
to high level ones.


5. In Summary...

In point 1, We have specified certain features our architecture must possess.
This is based on the assumption that they lead to the highly desired features
that we have outlined.

Point 3 shows us the way we intend to manage the complicated problem of
generating behaviour.

Clearly the idea of conception and self-enhancement is an innate objective in
the design of the architecture. The design must fulfil these two goals from
bottom up, as specified by point 4 in order to satisfy our criteria for
intelligence.

Point 5 specifies the general way in which stuff is processed in the system,
which is drawn from Connectionism.

Point 2 provides power to run the whole system through its 'drives'.

I STRESS THE IMPORTANCE OF THE SYSTEM TO MODIFY ITSELF TO ENHANCE ITS
ABILITIES. THIS IS AN IMPORTANT GOAL OF THE DESIGN.

MRI 3.3 What needs to be answered first, and what must be done to do so
-------------------------------------------------------------------------

1. In reference to point 4 of the previous section, we need to know how much
specifics we need to design and how much to let the system evolve.

2. How exactly do emotions enhance the normal S drive mechanism in biological
organisms? (do they? or how else did the Emotions evolve?)
Why do they enhance survival (by supplementing the S drive) in a way the S
drive alone cannot?

3. How do we design the drives into the architecture? Especially the E drive?

4. [THIS IS THE BIG QUESTION] How do the features specified in the previous
section's point 1 provide the advantages which we already see in biological
brains?

MRI 2

MRI 2.1 Consolidation Attempt
-----------------------------

We've gotten many ideas from many places. Lets try to consolidate them and find a good integration of all of them.


First is a list and synopsis of the key ideas we have picked from others and also those we created:

From Stephen Wolfram in "A New Kind of Science"
--------------------------------------------------
Complex Behaviour can arise from systems whose behaviour are based on simple rules. An example of such is Cellular Automata, which is composed of an array of elements whose behaviour is dictated by simple rules which are functions of the state of adjacent cells. This is reminiscent of neural networks, in which simple rules govern the firing of each neurone (at least this is so in their general behaviour)

An increase in the complexity of the rules does not lead to a corresponding increase in thhe complexity of the resulting behaviour of the system.

Perception and analysis is very much concerned with summarising the details of some raw input.
To accomplish this it is necessary to IDENTIFY PATTERNS, and COMPRESS DATA.

(pg. 627 paraphrased)
The power of human thought lies in its ability to store and quickly retreive a huge amount of information.

That allows us to make many connections between concepts. Making references to items in memory and forming new connections between them and new perceptions forms much of human thought.
(I ask, "what else is there to human thought?")


From the Connectionist Model of Cognition
-------------------------------------------
Concepts are represented as patterns of activity in neural networks.

These are present in material form as the strengths and weights of the synaptic connections between neurones.

Hebb's Rule is the method by which the connection strengths change.

If objects/concepts are represented graphically as paths, then the points of intersection of paths represents the relations between the two things...
e.g. John is a boy is depicted by the 'john' path intersecting the 'boy path'. The intersection is manifested in neural networks by the synaptic connections, and the strength of the connections means the strength of the relation.

From Subsumption
-----------------

We borrow the concept of layers of behaviour and agents each with its own simple, low-level task. Thus the complex behaviour we want an intelligent system to manifest can be split into simple specific tasks handled by simple systems.


Misc.
-----
Motor control is achieved by huge networks of pattern generators whose pattern output control the contraction of muscles that give rise to


Some of my own early ideas before those stated in MRI - 1
----------------------------------------------------------
Basic Principles of a good control system:

Decentralised operation - everything is everywhere and nowhere...in no particular place.

The architecture breeds its own intelligence. Hence the complexity of its behaviour comes from the running of the system itself...design of the architecture is only to enable as much of this complexity to evolve from the unsupervised operation of the system

Precision is not always necessary. The system should be tolerant of inprecision

The system should be adaptive. It should learn and optimise itself. (Consult
MRI 1.2)

Miscellaneous (Random and Unorganised) Ramblings on Intelligence #1

I have decided to put up three documents (MRI 1, 2 and 3)
These are thoughts over the past two months, eventually culminating in MRI 3, which sets the foundation for my pursuit in the field of artificial intelligence.

MRI stands for Miscellaneous Ramblings on Intelligence, and the first two MRIs appear disorganised and messy, which was the way they were first conceived...messy and illegible scratches on rough paper.

The language will not always make sense, because the meaning sometimes is too difficult for my linguistic skills to express in words. Sometimes this makes some sentences sound stupid. I understand some ideas are really stupid...haha hopefully they get refined in subsequent editions of MRI 3, where the ideas of MRIs 1 and 2 are consolidated, and more...



MRI 1.1 Some thoughts on the network
--------------------------------------

Note: Network here refers to a body of numerous interconnected elements


A network can exhibit a large variety of possible states, far greater than that of a digital computer.

(hmm okay, what exactly makes up the state is debatable. maybe this statement is not valid now on hindsight)



A network can possibly be a good model for a UNIVERSAL REPRESENTATION OF THE WORLD. This may provide much flexibility when having to deal with the complex dynamical system the world is.



We have a:
1. Representation problem (i.e. how to represent the inputs to the system within the system...the inputs being 'the WORLD')

2. Interaction of Representations (...to produce complex and interesting behaviour...Is thought solely or partially a product of the interaction of representations flowing around in the network?)



Perhaps by mathematical reasoning (or not), we could design an architecture that provides maximum flexibility and potential ('potential' as in huge possibilities in behaviour...or not?)




The network would come up with the THE MOST EFFICIENT WAY TO MAKE A REPRESENTATION so it can work on those REPRESENTATIONS in the simplest and most efficient way.
(hind-note: i.e. just like logarithms make multiplication simpler and fourier analysis reduce difficult integration to simple algebra...this is a way of self-optimisation of the network's own problem solving process, a sort of learning too. )


MRI 1.2 A little bit on Learning and Creativity
------------------------------------------------


Could a good and precise definition of learning make it easier for us to understand how STRENGTHENING THE CONNECTIONS BETWEEN NEURONES constitutes all there is to learning on the neurophysical level?
(OR is there more to strengthening connections in the brain's learning mechanism? perhaps on a higher level, probably built on from this lower level connection strengthening mechanism?)



What is learning? We attempt to define (not exhaustively):

A Set of Qualities (based on macro, behaviour-level layman interpretations of learning) which includes:

"Refinement" - adaptation, to do things better

"Inference" - taking in new knowledge and generating futher insights, making us wiser and "REFINING" the thought process for future demands

"Generalisation" - (definition feels incomplete) identifying underlying patterns which apply across a set of representations/things/mental objects.

Clearly, they all lead to refinement. Which means things get done faster, more efficiently, and less effort in the end. The end product of all learning.

Which leads mentally to emotional satisfaction and physically, survival in the material world.
Emotion is an internally generated purpose, drive.
Survival is an external purpose, drive.
Both drive learning.


(How about "UNDERSTANDING"? Where does it lie? between "INFERENCE" and "GENERALISATION"? What does it mean when we understand something?)




Creativity
Probably a process generated by a parallel mechanism...
- perhaps many approaches are all considered at once, and one of them wins? (sounds like 'neural darwinism')

- indirect/by-products of the mainstream cognitive process (i.e. conscious reasoning??)

- Or perhaps by-products of the interconnection of representations...More connections, more products, which are stored, and sometimes surfaced when needed. Hence more interaction, more creativity.




MRI 1.3 The Conscious Thought Process
-------------------------------------

Man's conscious thought process is a SERIAL one. We can only pay attentiona to one thing,
Perhaps this confers some ADVANTAGES, so the brain needs it. After all, we can do only so much at one time.

Perhaps this serial thought process is somewhat like the Turing machine.

But while the Turing machine is highly capable in executing algorithms, it isn't creative enough to come up with the algorithm itself...or is it? (Anyway the 'programming intelligence approach' hasn't worked well so we assume it doesn't work well)

This is how the human subconscious comes into the picture. It probably plays an important role in the creative process, in a way that was described above.

Perhaps the brain does things that way (i.e. running certain processes in parallel and others in serial) to optimise its ability to think.





MRI 1.4 The Problem
-------------------

Much of the world is probably about:
1. discovering the intersection of sets of objects
2. creating or destroying (in whole or in part) interactions between sets

This echoes the CONNECTIONIST model of human cognition.


We could like to create an extremely flexible and adaptable architecture for fast and efficient computation. Clearly a problem can be solved with a host of methods, some of which are simpler and faster and more efficient than others.

So can the architecture be designed to be flexible enough to learn to configure itself for the best possible problem-solving "SUB-ARCHITECTURE"?

[This coincides with the point on choosing the best mode of representation that would aid in finding the best way to solve a problem. That point was stated in MRI 1.1

Some slightly new developments

I said 'slightly new' because these 'developments' are largely confined to pen and paper for now, and some kilobytes on my harddisk.

Yesterday i obsoleted the UAV project i worked on in 2003-2004, archived the folder, and started a new UAV project. It would surely benefit from the ground previously broken by the old project.

The idea for the new project first came when i was waiting at the Tekong IMT centre for my turn to shoot. jotted an engine placement scheme in my notebook then left it alone for a while. Days ago i decided to make a UAV, and that scheme came to my mind. The flexibility and multiple redundancy associated with its VTOL capability is appealing to me, and i have just come up with a design i can start building. Two wings stacked as a cross, four engines on the ends of the wings. The engines are mounted on servo-controlled joints that allow a hemispherical movement space (hmm its difficult to explain, apologies for the bad use of words)

The X-Wing UAV (okay maybe its difficult to visualise) has aspects reminiscent of previous efforts by other pple...hovering discs or other skeletal crosses hovered by 4 engines, and the V-22 osprey's tiltrotor concept. I hope this 'rojak' design allows it to 'dance like a helicoptor, cruise like a plane'...if you understand what i mean.

Now to go work on the control system, which will be really difficult. Also, to muster the monetary resources to acquire my motors and batteries and structural materials and electronics parts.

Long way ahead. The success of this project is extremely crucial to myself.

I cannot fail.

gotta learn to like my new fate

oh well, the past few weekdays leading to the end of my Basic Section Leader Course (BSLC) have been heavy hearted. No it wasn't because i miss SISPEC or Tekong (oh no am I?) but the fear that i could be posted back to SISPEC again for the Advanced Section Leader Course (ASLC), which means i would have been an infantry specialist. I'm no action-man and charging around with six men under you isn't something i have aptitude for. The thought that i would be doing that makes me shudder, but in hindsight, that now seems nothing compared to what i may be going through...more on that later.

So as our sergeant read out each posting destination, followed by the names that would apply to, my heart pounded increasingly harder. For each name that was read that was not mine meant a greater chance of getting stuck on this island. Some lucky souls got sent to the airforce! Some were sent to the Unmanned Aerial Vehicle Training Centre (UAVTC), and i was disappointed i wasn't among them. I was posted to the school of armour instead. That left me quite affected, though i struggled to give myself a feeling of relief that i wasn't sent for ASLC. well its not that bad i told myself, tanks played a rather interesting role in the second world war, and besides its probably going to get more interesting than that regular infantry business. Yesterday i found out that things just got 'better'...or worse.

Of the Mortar Commander's Course, Tank Commander's Course and Armoured Infantry Section Leader's Course (AISL) , i was posted to the lattermost. What a blow as i saw the guy slide his ruler down the very long namelist to reach mine, and spotted the acronymn AISL. I looked at it a second time with disbelief, before the guy told me so. Then i was scolded by an old man who turned out to be the course warrant officer for smiling. i'm still quite bewildered, but i think he probably mistook my smiley nature for a lack of seriousness.

I felt a little better after having an early lunch with my friend, but emotions slid all the way while i spent the rest of the day at the library, then at Kinokuniya, and then at my aunt's place, where i managed to pick myself up and am still clinging precariously onto this fairly stable emotional state, which always threatens to give way to the depths of emotional hell.