Monday, January 21, 2019

Bugs

All computer programs of significant size have bugs in them. (As do the libraries, compilers, and hardware* they make use of.) I have spent much of the year so far trying to address some issues I ran into while assembling a large (A.s.a. H.) hierarchical case base from pieces learned by different robotic agents, both real and simulated, run at different times, in multiple environments, performing a variety of tasks.

Too many platforms, too many operating systems, too many languages, too many compilers....

* For example, with one of my Raspberry Pis the micro SD card must be plugged in just right. I've also commented previously on issues with LEGO plugs and the like.

Thursday, January 10, 2019

A.s.a. multitasking

We all know that multitasking while driving is a bad idea but most humans feel free to chat with companions while driving. Similarly, a portion of A.s.a. H.'s hierarchical network can be directing a robot to a recharging station* while another portion of the network may be extrapolating, interpolating, planning, etc. on some completely unrelated problems.

* Or transporting a solar array to a sunny location.

More redesign

I am redesigning A.s.a.'s robots trying to address the problem noted in my 13 Dec. 2017 blog. I want to put as many of the pain sensors (and lead wires) as I can inside the robot bodies. I may try to make greater use of my 3 Raspberry Pis while I’m at it.

The human brain has no pain receptors in it. A.s.a., however, can carry thermistors* and accelerometers inside its computer brain.

* Raspberry Pis, for instance, might experience overheating issues.

Tuesday, January 1, 2019

Wall confined plasmas

Classical theory greatly overestimates the confinement of beta > 1 plasmas. Intense particle and heat loss in actual experiments has made it difficult to reach (Tn)/(B*B) =1 in order to even study the wall confinement regime.

Robot controller

Many of my robots have been tethered to computers and/or power supplies. This constrains their operation somewhat. Because of its low cost (80 U.S. $) and large number of ports (39) I have bought an EZ-Robot EZ-B V4/2 WiFi robot controller to try out.* Due to limited funds (equipment) and lab work space I typically have to disassemble a robot or two  before I can build a different one.** The New Years break may give me the time that I need to do that.

* We may still need a tether to a power source for some experiments.
** My blog of 9 Nov. 2018 addresses this issue when only the processor needs to be changed out. Modular robot designs can also help.

An argument for alternative logics

Logic is about the formalization of sound reasoning. Since there are different ways of reasoning* there are different logics.

* Those few who might dispute this would argue for a single deductive reasoning. But that would require starting from a set of absolute and eternal truths and these are not available to us.

Medical AI use

Ulloa, et al, report (arXiv:1811.10553, 27 Nov. 2018) that a deep neural network can predict patient 1-year survival from echocardiograms significantly more accurately than trained human cardiologists.

The development of AI in general might take the form of creating specialist AIs like this and adding them to a growing society of agents.* Some of my work with A.s.a. H. has been along these lines. This is the pattern by which machines have taken over other tasks from humans and draft animals, etc. (i.e. automation)**

* Kiva logistics (warehouse) robots would also be an example.

** And see my blog of 20 Sept. 2018.

A danger in education research?

Learning is, in general, NP hard. But some problems are easy or easier. There may be a temptation to simply use education research to identify the easy subject matter (and methods) and then only teach those topics or those examples. This is likely to prove popular with the student community and administration. We might then end up only teaching classical (e.g. Newtonian) physics, for example, and neglect harder things like relativity and quantum mechanics. The difficult foreign language courses have disappeared from many colleges.

I see no problem with identifying and starting the students off with the easier material. I just want to be sure that we eventually cover important subject matter even when it is difficult.

Divided consciousness

Hilgard explored divided consciousness in humans. (Divided Consciousness, Wiley, 1986) A.s.a. H. thinks about its own thoughts when it extrapolates, interpolates, plans, etc. This, as well as things like attention, intention, and short term memory, is divided up across various levels in the A.s.a. hierarchy.*

My work on consciousness is part of a broad effort to understand and to give AIs adequate attention mechanisms. I’ve considered translating A.s.a. H.’s consciousness** into PROLOG and adding it to rule based expert systems. This would require fibring PROLOG with a temporal logic, however, so as to preserve the time order of various events/processes. (And assumes that the expert system has appropriate sensors, actuators, and operates in a world similar to the one A.s.a. H. was trained in.) I would also have to give the expert system similarity measures.

* Modeling across multiple levels of abstraction is important and was designed into A.s.a. H. from the beginning, T.K.A.S., vol. 109, number 3/4, page 159, 2006. See Stuart Russell in Ford's Architects of Intelligence, Packt, 2018, page 52. See also D. Estrada, Conscious Enactive Computation, arXiv:1812.02578, 7 Dec. 2018.

** Such as in my blog of 1 Jan. 2017.

Alternate realities

In his book Coherence in Thought and Action Paul Thagard explores the relative coherence of materialism, theism, and dualism. (MIT Press, 2000, especially page 119) While I agree with Thagard's general conclusion that materialism is more coherent than theism and dualism I would differ a bit on the specifics of his evidence and explanations. (pages 121-124) I also believe that materialism, dualism, and theism are each sufficiently coherent as to each constitute its own alternate reality, each being accepted by different groups of people.

Muller has argued that our various realities are emergent from a more fundamental first person state space. See arXiv 1712.01816 and 1712.01826. (Much in the way A.s.a. H. Creates its models of the world abstracting away from its first person sense impressions and actions.)

A.s.a. H. as a hierarchical genetic algorithm

One of the original learning methods I used on A.s.a. was the mutation of the strengths of the components of the various case vectors. This was employed on each level of the knowledge hierarchy.