Thursday, January 29, 2015

The importance of the curriculum for a learner

As a substitute for knowledge acquisition and engineering machine learning is frequently thought of as free. But you can't just release an AI (or a human infant) into the wild.  As John Andreae has observed with his AI "...run on its own.  It quickly runs out of memory...it is better for PP to be 'taught' by a teacher." (Associative Learning, Imperial College Press, 1998, pg 13)  With Asa H I find that what is taught and the order in which it is presented is quite important.

Friday, January 23, 2015

Man and machine

In recent years I believe that I am finding myself spending more of my time with machines (computers) and less time with humans.  I simply find machines to be more rational than people are.   This may be sad or it may be another sort of Turing test.

Wednesday, January 21, 2015

QNX

I have heard good things about the QNX operating system so I have ordered a Blackberry playbook tablet in order to try it out.

Free will and nonlinearity

Nonlinear descriptions of reality may be one of the origins of what we think of as free will.

For some problems (like robot motion planning) it is appropriate to explore multiple alternative solutions (e.g. alternative routes). Suppose an AI has learned to model some activity using a quadratic function.  For a given input condition it computes the (>1) roots of this model quadratic.  Even if the AI always picks a solution (root) in the same way (first found, smallest, randomly, etc.) it sees that another output (solution) would work too.  It sees itself as free to use either solution to its problem.

The AI is going to store and reuse some of its problem solutions.  As goals and external conditions change it may even start using other roots or choose from the available solutions (roots) in some different way.  A notion/concept of free will might develop from this.

With a society of Asa H agents I sometimes use an executive or router to assign tasks (or sent input) to one or more of the specialist agents.  I am looking to see if a concept like "free will" evolves in this executive.

Friday, January 16, 2015

Intelligent systems

In 2000 C. W. de Silva argued that an intelligent system would possess:

sensory perception
pattern recognition
learning and knowledge acquisition
inference from incomplete information
inference from qualitative or approximate information
ability to deal with unfamiliar situations
adaptability to new, yet related situations
inductive reasoning
common sense
display of emotions
inventiveness

and that the then "current generation of intelligent machines do not claim to have all these capabilities." (Intelligent Machines, CRC Press, 2000, pg 5)

I claim that my AI Asa H has now demonstrated all of these capabilities (to varying degrees).

Friday, January 9, 2015

Asa H value change

Intelligences may change their values over time.  Slavery was once accepted by humans, now it is not.  I studied value change during my work on Asa F (see Trans. Kan. Acad. Sci., 107, 1/2, 2004, pg 37).

During some experiments Asa H has done self monitoring, watching how things like memory size contribute to utility/value improvement.  (see my book Twelve Papers, pgs 15 & 16, available on my website www.robert-w-jones.com, book)  In this way Asa H defined and developed the concept "knowledge."

Starting with only two primary values, offspring (copies) and lifespan, a society of  Asa H agents has now promoted knowledge into this category and reported this to me.

(I frequently use a society of agents because groups make better decisions than individuals do for the reasons explained in my blog of 17 Aug. 2012.)

Wednesday, January 7, 2015

Je suis Charlie

See my blog of 25 Dec. 2014.

Simple simulation environment

Robot simulations are faster and more economical than real physical mobile robots. Any simulation can be thought of as 2 coupled Turing machines, one, agent p, representing the robot, and the other, environment q, representing the environment:


At any time step the robot sees an input vector x' and may receive rewards r.  It also produces an output vector y.

The environment at any time step receives an input vector y and generates a response vector x' r.  An especially simple environment is nothing more than a case-based reasoner or approximate look up table.

Asa H agents (serving as agent p) can be taught certain concepts/behaviors in such  a simulator.

Multiple memories again

Tulving and others have suggested that humans may have multiple memory systems.(see some possible examples in the figure below)  One might gain efficiency by using different representations in different memories acted on by different algorithms.  The simplest example might be 2 dimensional arrays to store visual information and 1 dimensional lists to store audio information.  The simplest implementation in Asa H might be to use 2 copies of Asa H on the lowest level in the hierarchy, one with NM=1 for audio input and one with NM set equal to the number of pixels in an image for visual input. (see my blog of 10 Feb 2011 for an example of simple code)  At some higher level in the Asa hierarchy the outputs of these two sensory modalities would then be combined together.(concatenated) Things like translation, rotation, and reflection operations would only be applied to the visual memory.  Things like time dilation would be applied to both.  More complex examples of the use (and usefulness) of multiple memories are also under study.


Tuesday, January 6, 2015

V-SIDO robot operating system

I downloaded a copy of Wataru Yoshizaki's V-SIDO robot operating system. (alpha version 0.42)  The documentation is in japanese but I managed to run the simulator by:

double click on       v-sido 0.42    file
double click on       bin
double click on       vsido             application
click on                  1: ########  box
click on                   OK              box
click on various green buttons on simulated robot moving it around

This worked fine. If a real robot is interfaced to your computer it is supposed to do whatever you make the simulated robot do. I don't own an ASRA C1 so I couldn't check that out.

Standard of living and quality of life

Standard of living/quality of life, should be a vector quantity having components like:  health, safety, autonomy, housing, resources, nutrition, education, employment, influence, etc.  It should not be collapsed down to a scalar. You can only be sure of an improvement if all of the vector components have improved.(or some stayed unchanged while all others improved)

Thursday, January 1, 2015

The relationship between my AI Asa H and theories of the human mind and brain

Asa H is built out of case-based reasoners.  It is Roger Schank's view that "Case-based reasoning is the essence of how human reasoning works." (Case-based planning, K.J. Hammond, Academic Press, 1989, pg xiii)
Asa H operates on patterns in much the same way that ART networks do.  "ART was introduced as a theory of human cognitive information processing." by S. Grossberg. (Brain-mind machinery, G. Ng, World Scientific, 2009, pg 146)
The Asa H hierarchy operates in about the same way as does J. Hawkins' HTM.  HTM is Hawkins' model of the human neocortex. (On Intelligence, J. Hawkins, Times Books, 2004)
In Asa H clustering modules perform categorization.  This appears to occur in neocortical layers II and III  in the human brain.(Rodrigues, et al, J. Cog. Neurosci., 16, 856, 2004)
D. Hofstadter argues that "Analogy is the core of all thinking." (Surfaces and essences, Basic Books, 2013) Analogy is one of Asa H's basic learning/extrapolation algorithms.
Granger, Rodriguez, et al, identify the brain's computational instruction set as consisting of: sequence completion, hierarchical clustering, retrieval trees, hash coding, compression, time dilation, and reinforcement learning. (AAAI technical report FS-04-01, 2004, pg36) Asa H makes use of most or  all of these.

AI vs traditional programming, automatic programming

Most of the programming examples we teach in an introductory programming course have all or most of their functions built in before run time. Things like accounting programs, databases, inventory programs, etc. But it's possible for programs to acquire (some of) their functions during the run, and completely without human intervention.  This is, of course, a matter of degree. It's like heredity vs environment with humans.

Those functions that are acquired during run  time can come into the system with the data input stream (including any performance evaluation inputs/utility measures).  These functions may be spatial-temporal patterns seen in nature and learned by the program. Self-organizing systems would be a typical example of such a program as would be my Asa H, neural networks, and many other AI systems.