Tuesday, June 28, 2016

EV3 simulator

Things like the arduino simulator have been useful in separating the design and debugging of software from the design and debugging of robot hardware.  EV3 simulators are now becoming available.  I have just got the TRIK Studio 3.1.3  EV3 (and NXT 2.0) simulator up and running.  The biggest issue with it is the fact that the documentation is in Russian.

Saturday, June 25, 2016

Observation versus experiment

I have emphasized the importance of AIs being able to act in/on the world. (For example in my blog of 23 March 2015) They need to be able to experiment as well as observe.

If one simply observes the operation of a Jacob's ladder two alternative theories of operation come to mind.  Hot air rising between the electrodes may be lifting the arc or, alternatively, a Lorentz force may exist due to the current carried through the arc and the magnetic field produced by the arc and electrode loop. The Lorentz force would then propel the arc upward. 

As an experimenter one can simply tip the apparatus on its side and see that the arc no longer travels alone the electrodes, the hot air explanation is the better one. The agent/experimenter's ability to act on the system being studied is quite helpful.

Thursday, June 23, 2016


Because of the complexity and expense involved I have avoided humanoid robots (blog of 18 Feb. 2016).  But to give Asa H the concept of "walk"/"walking" I have located a few sensors on a small Lego NXT walker.

Tuesday, June 21, 2016

Student projects

How do you involve students in your research? You have to find a manageable chunk that you believe will lead to a clear and useful (publishable?) result in a reasonable length of time.  For a Ph.D. This needs to be doable in a couple years and it needs to be original research.  For an MSc perhaps you have only about a year.  For an undergrad project, just a few months.

I sure couldn't put students on my project looking for concepts with which to reconceptualize reality.  With that work you might get a single concept, like non-Markovian models, after a year or more of effort.  You can't just say, maybe we'll discover something interesting in a year or two.

As a consequence of capitalism business practices are being injected into every aspect of life. The idea of a masters thesis contract is a case in point.  Science is not business. One can not predict what will be discovered or when.

I suspect this all distorts the scope and direction of scientific research.

Saturday, June 18, 2016

Ultimate reality, on various scales

We don't have direct access to ultimate reality, we have only our sense impressions, our sensitivity to light, sound, pressure, temperature, and certain chemicals. (Some other creatures have sensitivity to electric and magnetic fields.) Like Mariam Thalos (Without Hierarchy, OUP, 2013) I do not believe that ultimate reality is reducible solely to the microworld, be it strings, or branes, or quantum fields, or Hilbert spaces, ...  The macroworld is also ultimate, be it a multiverse, higher dimensional, or whatever. There are also things like mind, thought, and consciousness which are patterns/processes. What these are patterns of, or processes in, is of less importance.  Asa's thoughts can be patterns of activity in an electronic computer or in an optical computer.  A computer, in turn, can be assembled out of matter in our world or equally well out of gliders in a universe like Conway's game of life.

Tuesday, June 14, 2016

Scientific pluralism, probability and statistics

There are a number of different theories of probability: objectivist, subjectivist, frequentist, Bayesian, etc.  (see, for example, AI: a Modern Approach, 3rd edition, 2010, pg 491) Statistical inferences are then made based upon a variety of competing approaches, each with its own different strengths and weaknesses. (See, for example, S. N. Goodman, Science, vol. 352, 2016, pg 1180)  In general one can not make claims based upon a single estimation of statistical significance, be it Fisher's P value, Bayes factors, or the like. Rather, one needs a pluralistic approach to value/assessment. (In a society of Asa H agents I have used various different value functions/networks.  See, for example, with Asa H light, my blogs of  10 Feb. and  19 Feb. 2011)

Saturday, June 11, 2016


Asa H can be taught names for the concepts it learns, for example:
Collision=(sense near, bump, decelerate) can be expanded to (taught):
Collision=(sense near, bump, decelerate, sound "collision")

Artificial neural networks, on the other hand, are frequently subsymbolic.

How many of the concepts (case vectors) that Asa learns should be named (symbolic)?

Going in the other direction Theodore Sider has suggested that complex linguistic entities be constructed as sequences or tree-structures of linguistic atoms (words). (Writing the Book of the World, OUP, 2011, page 295) This is exactly what Asa H creates (learns). We would certainly not want to assign names (words) to all of these larger scale case vectors. I.e., Vocabulary choice is required at this point.

Wednesday, June 8, 2016

Patom theory

John Ball's Patom theory (Speaking Artificial Intelligence, ComputerWorld, 2015) is quite similar to my A.s.a. H but has not yet been developed into code.  When it has been I will be interested in seeing just what design choices have been made and how it performs.

Tuesday, June 7, 2016

Downward activation in Asa H

In Asa H predictions and output actions are the result of downward (backward, top-down) activation.  They involve a flow of activation from upper levels in the hierarchical memory to lower levels.  It is also possible for an active case to send activity downward to all of its vector components.  This additional activity could then influence what other patterns (cases) might be or become active.

Friday, June 3, 2016

Lego NXT stigmergy

A swarm of Lego NXT colored brick (or beacon) seeking robots and brick (or beacon) dispensing robots can employ stigmergy to self-organize.

Wednesday, June 1, 2016

What is simple?

There are those who believe that our universe is not as simple as an empty one would be and that an explanation is needed as to why our universe is as complex as it is. There is more than one issue here but firstly, just what constitutes simplicity? Here are some possibilities:

1. What is more easily learned.  (But by which learning algorithm(s)?)
2. Shortest. (But in what language? Which representation?)
3. Have the greatest symmetry. (In which geometry? And which object properties are to remain unchanged by the transforms?)
4. Information-theoretic measures. (Which regularities should be counted?)

More likely what people mean by simplicity is a vector quantity again, a cluster of components. (Just like the concepts Asa H learns. For Asa most concepts are vectors.) And people won't even agree on those components.