deep learning of possible logical courses of action -
using digits
April 17 2018 | by Korvin AG |
The applications of
commercial AI techs are now becoming pretty much widespread
and with the tremendous capacity boom regarding processing
speed, storage size and sensory sensitiveness these are
yielding more and more for the industry. But every time it
comes to the age-old question, whether "could machines one
day think as humans do?" there is no clear answer yet.
Partly it is because it is yet to be
modeled how actually humans thinking in general and how the
thinking of given individuals differ from each other. And
believe me, that ain't gonna be a model to be digested
easily.
I started to code a very
simple logical course of action selector a while back, and I
found that even in the field of the machine
learning/artificial intelligence field the core design
principle will remain the classic
'60s rule:
"Keep it Simple, Stupid ! "
I started with
researching the possibilities regarding a car that is
traveling among the non-AI cars that are going in every
possible directions. What I found is that the flow of
the traffic, when seen "from above" is definitely
following the rules of hydrodynamics, but it is down to
sequences of haphazard events when You see it from the
perspective of the individual driver.
The most simple solution
would be of course to use the most out of the AI-driven
car's sensors. Watch out for the signs, the surroundings,
the other vehicles - and have these permanently synchronized
with the on-board (or on-line) geospatial/geological
informations database(s).
But this approach is in
line somehow with the idea to have someone drive on all the
roads of the Earth in order to get a driver's license - just
to learn what is possible when driving. So, here is where
the human thinking vs. artificial intelligence thing kicks
in.
To cut a long story
short, what I found was that the human decision is not
necessarily based on sensing ("what one sees", "what
one hears", etc.) and comparison with the full knowledge
("what one knows") - but on the possible ensuing logical
course of action. (Even though the logic involved
might not be 'logical' in some instances.)
For this reason I
started to collect a set of possible logical courses of
action, that can be taught to AI-driven systems. In order to
keep it learn-able and simple, these are all have to be
represented as processes in order to keep it possible to
use them in both feedforward and recurrent digital neural
networks.
In the end I was able to
use 54 such logical courses of action which finally went
into an anti-fraud fintech application. The number of these
now reached the 400 mark, but for this writing I will only
use the most simple and convincing set: the digits based on
the Arabic numerals.
For the purpose of this article letters are representing breakpoints (which can be translated also as interrupts or incidents) in the process of going from one place to another in an autonomous way. The arrows are representing the direction of the process, which equates to the heading of a self-driving vehicle.
Let's start with the
number 1. In my model it is rather like the
capital letter 'I'. It represents the uninterrupted
process, going from point 'A' to point 'B'.
The number 2
represents the event when in order to go forward and
towards the destination of the self-driving vehicle, it
has to get around an obstacle. After circumventing it,
the self-driving vehicle rearranges itself at point 'B'
and begin its uninterrupted journey towards its
destination, which is in this case the point 'C'.
The number 3
represents an event when the self-driving vehicle has to
get around an obstacle and immediately after this is
done, it has to do so once more again. At point 'B' it
rearranges itself and decides to circumvent the second
obstacle. At point 'C' it rearranges itself and can
progress towards its initial destination. See below for
an example: the DeLorean does it!
There are two
different uses of the numeral '4': the open top numeral
4 and the closed top numeral 4. The open top numeral
4 represents the logical course of action, when
two processes ('A'-->'B') and ('C'-->'D') are
intersecting each other at one point and there remains a
possibility that these processes, sometimes in the
future, will intersect each other once more. This is a
very prevalent event in the case of highway and
speedway traffic for example. The closed top numeral
4 represents the case when the self-driving
vehicle has to rearrange itself several times, because
the obstacle can not be passed by getting around
it. A nice example is when the self-driving vehicle
enters a dead-end street (or one that is blocked
temporarily). In this case it has to conduct a Y-turn,
and in doing so it intersects its previous track.
The current form of
the numeral 5, however, is replaced by a sign
that resembles a numeral 2 turned upside down. This
logical course of action too is the opposite of what we
seen in the case of the numeral 2: the self-driving
vehicle is progressing in an uninterrupted way, but at
point 'B' it detects an obstacle and decides to get
around it in order to reach its destination. At point
'C' it rearranges itself and follows the original travel
plan. A single event regarding this logical course
of action with the DeLorean:
The numeral 6
could represent many things, but the basic event is when
the destination is not accessible to the self-driving
vehicle. In order to deliver the cargo it has to decide
that instead of circling around the blocked off endpoint
(that is breaking the logical loop) it would stop and
drop off the cargo.
The numeral 7
I used here is the handwritten form of the number. It
represents the case when there are two parallel
processes, one of these being an uninterrupted one
('D'-->'E'). The other process ('A'-->'C') is
interrupted at a breakpoint (point 'B') which forces
this process to intersect the 'D'-->'E' track of the
process. This is what is happening when the self-driving
vehicle is overtaking another one, during which event it
has to use the oncoming lane. This means that at point
'B' the self-driving vehicle starts a maneuver which
intersects the track of the opposing traffic. This is a
one-time possibility and the AI must use its
capabilities to avert an accident by knowing that it
could occur. A simple scenario follows with the
DeLorean:
The numeral 8 is a representation of any
logical loop caused by malfunctions or an inability to
find a way out from a situation. Main characteristics
are that it continuously performing the same moves
(running the same sets of algorithms) and it crosses its
own tracks over and over again at approximately the same
place (after the same sequences). This could be useful
to train an AI/ML-driven system to be able to escape
logical loops. And this in itself adds a lot to the
capabilities of an AI, since it makes harder to spoof
it.
The numeral 9
is the opposite of the numeral 6 and is used here to
train the self-driving vehicle to understand the need of
other vehicles for purposely circling around a
destination. This is what we do all the time when we are
circling around in our cars searching for a parking
space. We might circle many times around hoping that
someone will leave its place. And this logical course of
action teaches the AI that there is a relationship
between 'supply' (e.g. number of parking spaces for
example) and 'demand' (the number of cars circling
around). And of course it makes the AI able to
differentiate between logical loops and temporary loops.
The numeral 0
is the way how the AI could relate to its own existence.
This logical course of action is the circulation that is
the normal way of operation for a self-driving vehicle.
This could involve things like departing the zero point,
arriving at the endpoint, refueling, performing
technical self-checks, departing the zero point, etc.
etc.
Of course this is a
greatly simplified abstract of the conception, but I
hope that some of you will find it useful :)
Thanks for reading and please feel free
to share:
Korvin AG
( If You want to contact me regarding this article or for any other reason, You could do so by sending an email to info@korvinag.com )
This text is released under Public
Domain, under a Creative
Commons Attribution 3.0 Unported License. Feel
free to use, reuse, modify, etc. Just don't forget to give
appropriate credit.
