Saturday, July 30, 2011

A feature discovery heuristic

Search for patterns localized in space and/or time.  Temporal localization can be accomplished using the temporal sequence heirarchy built into Asa H (see my Trans. of the Kansas Academy of Science 2006 paper on Asa H). Temporal patterns are of various durations; i.e., of various scales.  Higher levels in the hierarchy record longer duration patterns. Spatial localization can be accomplished by deviding the visual field into a grid of cells.  Each cell contains a set of points in 2-D.  Patterns are learned over the points in each of these cells.  The cells can be overlapped and cells can be of various different sizes.

Advanced life forms

I have argued in favor of mechanical life (see my blog of 19 Oct. 2010). This can be engineered rather than an evolutionary kludge. (There would be no vestigial organs, etc.) In addition to repairing components or "melting them down," the fusion torch (Eastlund and Gough, WASH 1132, 15 May 1969) would create a fully closed matter cycle; mechanical life need not input new matter ("food") nor output any waste matter. In addition to chemical energy mechanical life could make use of nuclear (or other) energy, performing larger amounts of work than is possible with present-day ("natural") biological organisms. Mechanical life forms would be super organisms.

If space aliens are capable of traveling to earth we might expect them to be of this sort. (Because of the high energy demands for relativistic travel I assume these should be as small/light as possible.)

Wednesday, July 20, 2011

All knowledge is uncertain

The majority of published scientific papers are wrong (J. P. A. Ioannidis, PLoS Med, 2(8), 30 Aug 2005).  We must just do the best that we can.  (Science is still better than business where the majority of companies fail in 3 to 5 years and 90% of companies fail in 10 years.)

Sunday, July 17, 2011

Soccer

Soccer is too low scoring.  I suggest that the goal be made bigger. Or will there be too many corner kicks?

Saturday, July 16, 2011

Vector value in the human nervous system

I have argued that utility/value should be a vector quantity.  (My post of 19 Feb. 2011, my website,
http://www.robert-w-jones.com/, philosopher, axiology, and references therein) Schultz (J. of Neurophysiology, 80: 1-27, 1998) notes that the responses of most dopaminergic neurons are similar, suggesting that this system carries a scalar signal.  Schultz and Romo, however, (J. of Neurophysiology, 63: 607-624, 1990)
report that different goups of neurons display both excitatory and inhibitory responses to the very same situation.  This would suggest vector reward signaling. (see the 2003 PhD thesis of Nathaniel Daw, Carnegie Mellon school of computer science)

Roomba

We own 2 roombas (1 "roomba" model 4105 and 1 "roomba scheduler" model 4230).  I had thought to hack the roomba (Hacking Roomba, T. Kurt, Wiley, 2006) to produce and study an embodied agent. I have experimented with the roomba as it comes "out of the box."  I have a bookshelf thats a bit over a foot deep and about the same distance from a corner of the room.  Roomba gets stuck in that three sided box.  Roomba also gets stuck half under one cabinet, my rocker, and tries to climb up the base of a floor lamp (placed near the wall and a drape). As it is I don't think roomba will work even in a "slightly" cluttered room. (Most kid's rooms would be much more cluttered than that.) I suspect that if I hack roomba and control it with Asa H I might be able to produce behaviors that would fix these issues.

Thursday, July 14, 2011

Sixth sense

There are bacteria that can sense magnetic fields and sharks can sense electric fields.  These are real sixth and seventh senses.  But humans don't have them.

Friday, July 8, 2011

Free will in men and machines

There are a number of different theories of free will (R. Jones, 2002 annual meeting of the Kansas Academy of Science, Fort Hays State University):
Theory #1: Free will is a self-contradictory notion (Richard Double).
Theory #2: Free will, though conceivable, doesn't exist (Marvin Minsky).
#3: Some behavior randomization is used in order to make us unpredictable to adversaries/predators/competitors.
#4: "Free will" is simply a denial that CERTAIN SPECIFIC sorts of influences were operating.
#5: Cognitive system nonlinearity allows multiple responses to identical external stimuli.
#6: Decisions are based on history/path dependent (and approximate) processes.
#7: Different decisions can result under identical conditions if the agent deliberates for different lengths of time. ("Delay libertarian theory." "Anytime algorithms.")
#8: Quantum mechanics makes it impossible for the same decision to be made again under exactly identical conditions.
#9: An agent (using genetic algorithm or other unpatterned generator) creates rules for how it should decide on actions and tests them in its daily life. Such a random algorithm could just as well have created different rules. (We are most free when we are creative, doing something truly original/unique.)
#10: Everett-like multiple worlds, your decision determines which world you enter. (A strong "free will" requires extreme assumptions about physics.)
With our current knowledge I do not see how a strong free will (like that assumed in religion and the justice system) is possible.  On the other hand, a weak free will (like #3 above, for instance) is certainly reasonable.
I think what passes for free will in humans is a combination of a number of processes like #4, #5, #6, #7, #8, and #9 above.  Complex machines would be subject to the same principles.

Wednesday, July 6, 2011

Media Bias

In watching the Casey Anthony and DSK coverage I note that the US media has a strong pro-prosecution bias. The more rightwing media is the worst. People think they have to have someone to hold responsible.  Do people really believe in innocent until proven guilty?

Friday, July 1, 2011

The Cycle of Theory and Experiment

Science consists of a cycle of theory and experiment.  Experiments are needed to "keep us honest."  I have criticized Hutter's AIXI (Universal Artificial Intelligence, Springer, 2005) for offering no executables. (Joel Veness, Hutter, et al, now have provided pseudocode in "A Monte-Carlo AIXI Approximation", Dec. 2010 and C++ code on Veness' website, jveness.info/)
But it's hard to get something more than qualitative results from AI experiments.  In my Asa H I have a choice of many clustering algorithms (or a combination of several of them), various similarity measures, a variety of feature extraction algorithms, and multiple extrapolation/learning algorithms.  I can also adjust the amount of processing time devoted to learning and to each kind of learning.  There are also a number of thresholds/learning rates/free parameters to set.  I have found that in these ways AI experimentation is much harder than the pure physics experiments I did 20 years ago.