Tuesday, September 28, 2010

Stopping undersea oil leaks

When the deepwater horizon incident occurred I sent the following suggestion to British Petroleum and Washington:

"An explosion near the well head on the ocean floor could remove the blowout preventer and shear the pipe off at the sea bed.  A second explosion to one side but near the exposed pipe could push soil over the well and seal it.  It might also be possible to crush/crimp the pipe."

Monday, September 27, 2010

Problems with electronic submissions

Electronic submissions for publication have their problems.  In the old days when you sent in a paper copy of an abstract they might occasionally cut off an edge when it went under the camera.  I will report on my recent experiences with the Bulletin of the American Physical Society because they have been the worst but I have had similar problems with other journals and conferences.

My last submission to BAPS had a % symbol dropped.  This did not occur in the echo to my electronic submission.  I was using the American Physical Society's own software (Latex).

The submission to BAPS just prior to that one had no errors that I noticed but the one just before that had 5 extra characters inserted in various locations throughout the abstract.  In past years I have had as many as 3 or 4 words dropped in an abstract.  Sometimes the sense was lost completely.

Sunday, September 26, 2010

Scientific Pluralism

Knowledge is of an approximate character.  Our formalisms abstract and simplify.  Each theory is an idealization, often times approximating in its own DIFFERENT ways, each offering somewhat different coverage of the domain of interest.  Having MULTIPLE overlapping theories of a field is then better than having just one.
(www.robert-w-jones.com, Philosopher, changing what science is and how its done)

Throughout my career I have followed this "scientific pluralism" by trying to work on both sides of the various important questions: (for example)

magnetic confinement of plasmas     -versus-     inertial confinement

low beta plasmas     -versus-     wall confined plasmas (beta>1)

open magnetic traps     -versus-     closed magnetic confinement

adiabatic traps (magnetic mirrors)     -versus-     nonadiabatic traps (cusps)

neural networks     -versus-     g.o.f.a.i. (good old fashioned AI)

scruffy AI     -versus-     neat AI

dualism     -versus-     monism

rockets     -versus-     tethers

rockets     -versus-     space drives

I also try to do theoretical, experimental, AND computational work.

Tuesday, September 21, 2010

Value system for a mind

Logic of the mind

While present day computers execute classical propositional logic using operators like AND, OR, NOT, and IMPLIES (and thence NAND, NOR, EXOR, ...) a mind would require a fibring with, at least, temporal logic with operators like SINCE, S(A,B), UNTIL, U(A,B) (and thence, F, "in the future", P, "in the past", ...) as well as spatial logic with operators like "in front of", "behind", etc. and a non-monotonic consequence relation. (And probably the logic should be fuzzified.)

In addition to the fibred logic a mind would also need a value function/system.

Friday, September 17, 2010

Computer simulations versus "real" experiments

In order to pursue online instruction there is a desire to use computer simulations in place of "real"
experiments.  We even hear those who claim simulations are "just as good as" real experiments.

There are various important ways in which simulations differ from real experiments but the bottom
line is:

simulations only contain the physics that you understand and were able to program in,
real experiments also contain all the physics you haven't thought of, don't understand,
and haven't included.

Thursday, September 16, 2010

Growing your research program

Most work is done by groups.  You likely begin your career working on a PIECE of a project in some
SUBFIELD.  An adviser may have given you your task and will know it is doable with the resources at
hand, is worth doing, and can help with debugging/troubleshooting if you run into trouble.

After a few such efforts and a lot of reading of the literature you will be able to choose and perform
experiments relatively independently.  Your need for the help of others will decrease.

You should first work to become an expert in your particular subfield.  This may take typically  perhaps
5 years. You should try to maintain frequent contact with coworkers both at your institution and
internationally.

One would slowly expand his study to encompass much of his entire subfield and then even explore
other portions of the parent field (but being careful not to spread yourself too thin).

An ideal job would give you the time and freedom to do these things as well as laboratory facilities,
libraries, funding, etc.  (Some time may be required just to secure such a position!)

One will typically keep files on his own work organized by project, subfield, and field.  Similar
reference files would be kept containing the work of others organized following the same system.

Getting started in research

Many graduate students feel that they are doing all the work and their adviser is doing little while putting his/her name on any publications that result.  It is useful to list some of the contributions that an adviser makes (beyond building the laboratory, taking any data himself, writing funding proposals, etc.).

An adviser typically identifies what work remains to be done in the given subfield one is working in.

An adviser identifies what problems can be attempted TODAY with the resources available LOCALLY.

The adviser can often times break the problem into pieces and perhaps even parcel these out to a group
of researchers (perhaps students).

The adviser often times is able to troubleshoot when something goes wrong.

The adviser/senior researcher may know what work is most worthwhile ("publishable").

The senior researcher likely has an extensive file of important/useful references that will help with a given
project.

Much of the adviser's contribution may be years worth of work performed long before the student came
on the scene.

In my case I must thank professors Milos Seidl and Wayne Carr for the start they gave me.

Intelligences

1.  fixed database/rulebase
     no learning
     no value function

     An example would be an expert system.

2.  learning
     no value function

     An example would be a typical neural network.

3.  learning between generations only
     value function used

     An example would be certain genetic algorithms.

4.  learning during the lifetime
     value function used

     An example would be a human (but humans have limited rationality, see, for instance
                                                      Predictably Irrational, Dan Ariely, Harper Collins,
                                                      2008, and a flawed value system founded on a few
                                                      simple drives and aversions).

     Asa H may be a better example (having a better value system and being more rational
                                                      www.robert-w-jones.com, inventor, A.s.a.).

5.  collective intelligences

     An example might be a society (it has been known for 100 years that groups can make
                                                     better decisions than an individual can).

     Another example may be a collection of Asa H agents.                    

Monday, September 13, 2010

Minimum private property

In distinguishing socialism from communism the issue of the amount of allowed private property comes up.
The amount of private property needed would surely depend upon where a person lives, what job they do, etc.  I would think we would all need a few of our own clothes, toiletries like comb, tooth brush, razor, etc.  We might need a watch and cellphone.  Until recently we would need files and notes and specialist's books (in my case for my research and the textbooks I teach from).  With the advent of e-readers like kindle we might instead be able to access books and papers not found in our local libraries.  In certain remote locations one might need private transport.  We might also need glasses (as I do), hearing aids, etc.  We would all have a few personal items like pictures of loved ones, etc.

What constitutes "core AI?"

Core AI is supposed to be the most basic most fundamental areas of AI.  This is not quite the same as the most important areas of AI.  It probably is what should be in any good AI textbook.  Certainly I would include in core AI:

1. search
2. learning
3. memory

as well as something many other researchers would omit:

4. values/utility/fitness

Most people would also throw in:

5. logic(s)
6. representation

I think a good case can be made for adding:

7. pattern recognition
8. complexity
9. heuristics
10. compression
11. statistics

Depending on how broad you want your coverage to be one would also consider (in no special order):

12. natural language
13. neural networks
14. planning
15. consciousness
16. theorem proving
17. parallel computing
18. feature extraction
19. classifiers
20 modularity
21. architectures

Friday, September 10, 2010

How much we know

We have the impression that mankind knows a great deal.  This impression may come from the fact that for 12 to perhaps 24 years we are taught mankind's collected knowledge in our schools.  Year after year we hear of mankind's accomplishments.  But we typically only hear of the things mankind does understand.  As you might expect our teachers (wisely) have only a little to say of those matters we don't understand.  So our impression is biased and may be wrong.

Thursday, September 9, 2010

AI Subfields

Large difficult problems are frequently solved by first breaking them up into a set of interrelated smaller problems.  The AI subfields can be a set of such smaller problems into which AI is decomposed.  It is also useful to have a set of subfields that you can go to in order to find methods, algorithms, etc. that can be helpful in your AI work.  No such list is ever complete or unique but here is one I use:

1. weak methods
2. search
3. rules
4. semantic nets
5. logic/deduction
6. heuristics
7. discovery/creativity/induction
8. natural language
9. neural networks
10. distributed AI/collective intelligence
11. robotics/embodiment
12. compression
13. automata/state machines
14. statistics
15. Bayesian statistics
16. planning/scheduling
17. case-based reasoning/memory-based reasoning
18. blackboard systems
19. nonstandard logics (spatial logics, temporal logics, higher order logics, multivalued logics, etc.)
20. representations
21. consciousness
22. learning/data mining
23. theorem proving
24. automatic programming
25. genetic programming
26. qualitative reasoning
27. constraint-based reasoning
28. agents
29. fuzzy logic
30. diagrammatic reasoning (and spatial logic)
31. model-based reasoning
32. emotion
33. ontology
34. quantum computing
35. analogy
36. parallel computing
37. pattern recognition/comparison
38. causality
39. deductive databases
40. language of thought
41. artificial life
42. philosophy of AI and mind
43. innateness/instinct
44. AI languages
45. memory/databases
46. decision theory
47. cognitive science
48. control system theory
49. digital electronics/hardware
50. dynamical systems
51. self-organizing systems
52. perception/vision/image manipulation (and spatial logic)
53. architectures
54. complexity theory
55. emergence
56. brain modeling
57. modularity
58. hybrid AI
59. optimization
60. goal-oriented systems
61. feature extraction/detection
62. utility/values/fitness/progress
63. multivariate function approximation
64. formal grammars and languages
65. theory of computation
66. classifiers/concept formation
67. theory of problem solving
68. artificial immune systems
69. curriculum for learner
70. speech recognition
71. theory of argumentation/informal logic
72. common sense reasoning
73. coherence/consistency
74. relevance/sensitivity analysis
75. semiotics
76. machine translation
77. pattern theory
78. operations research
79. game theory
80. automation
81. behaviorism
82. knowledge engineering
83. semantic web
84. sorting/typology/taxonomy
85. extrapolation/forecasting/interpolation/generalization
86. cooperation theory
87. systems theory
88. semantic computing
89. exploratory programming
90. specialization/decomposition

There is, of course, lots of overlap in these.  Some are certainly more important than others but each might be
a source of help/information/inspiration. (My AI research library is organized into these 89 clusters of books/papers, plus some miscellaneous, textbooks, etc.)