Slide 1

A Thought on Intelligent Computing Approaches to
Challenging Problems
Toshinori Munakata
宗像俊則
email [email protected]
URL
http://cis.csuohio.edu/~munakata/
AI
Techniques
Apply
Challenging
Problems
AI Techniques
Functional type
Pattern recognition
Optimization
...
Specific technique
Evolutionary computing (GA)
Neural networks
...
Challenging Problems
There are many, probably thousands, but we consider:
o Analysis of Earthquake Prediction
o Data Analysis for Volcanic Eruption Prediction
o Analysis of Photosynthesis (光合成)
o Tornado (竜巻) Prediction
o Thermonuclear Fusion (熱核融合) Research
Each of these problems is challenging (which means hard),
but if solved, is possibly worth a Nobel prize.
Recent successful AI techniques reported
1. Neil Savage. Automatic Scientific Discovery, CACM, May
2012, p. 9.
2. Waltz, D. and Buchanan, B.G. Automating science, Science
324, 5923, April 3, 2009.
Basically an application of evolutionary computing (GA)
Neil Savage. Automatic Scientific Discovery, CACM, May 2012,
p. 9.
“The team fed their algorithm with experimental data about yeast
metabolism, along with theoretical models in the form of sets of
equations that could fit the data. The team seeded the program
with approximately 1,000 equations, all of which had the correct
mathematical syntax but were otherwise random.”
Basically an application of evolutionary computing (GA)
o Analysis of Earthquake Prediction
Every earthquake is unique and very nonlinear.
This means that trying to derive general and universal
techniques applicable to all cases is probably not productive.
In Munakata’s opinion, researchers give up too soon.
All earthquakes
Predictable earthquakes
Animal behavior
Electro-magnetic observation
(e.g., 服部 克巳 千葉大学理学部 地球科学科 教授 – 電離層異常)
...
__________________________________________
Phase I. Collection of past observed physical and chemical
measurement data and existence and non-existence of earthquakes.
Phase II. Application of AI techniques to analyze the observed data.
o Data Analysis for Volcanic Eruption Prediction
e.g., Mt. Fuji
Should be easier than earthquake prediction
(narrower geographical area, narrower focus on types of
measurement)
Types of measurement (Wikipedia)
2.1 Seismic Waves (Seismicity)
2.1.1 General principles of volcano seismology
2.1.2 Seismic case studies
2.1.3 Iceberg tremors
2.2 Gas emissions
2.3 Ground deformation
2.4 Thermal monitoring
2.5 Hydrology
2.6 Remote Sensing
2.7 Mass movements and mass failures
Collect past data for the above measurements (inputs) and
subsequent volcanic eruptions with their intensities (outputs).
Determine mapping from the inputs to the outputs by AI techniques,
e.g., ID3, backprop, etc.
Once mapping is determined, it can be used for real-time prediction.
o Analysis of Photosynthesis (光合成)
A “breakthrough” was reported in “Plant Life’s Boxy Heart,” Science,
334, 1630, December 23, 2011.
Our basic idea:
Collect current theories and equations to represent photosynthesis
process.
Represent these equations with parameters (if necessary, rewrite or
modify the equations).
Generate many (e.g., 1000) different versions of these equations by e.g.,
slightly modifying parameter values.
Apply AI techniques (e.g., evolutionary computing) to determine a best
set of parameter values. Repeat the last two steps if necessary. )
o Tornado (竜巻) Prediction
This is a major problem in the U.S.
Tornado prediction models do exist and they give decent forecast for
tornadoes. We want to improve the current forecast (e.g., earlier
warning, more accurate area description) by employing more powerful
computers, better algorithms (e.g., AI-based, e.g., parameter
optimization), and faster communication.
For recent hurricane forecast two weather models were employed: the
US and European. The latter performed much better. The much higher
computing power used for the European model is credited for the
difference. e.g., in the European model, much finer mesh points were
considered for calculation.
It is conceivable to transmit weather data from the U.S. to Japan,
compute prediction by a supercomputer (such as 京 ), and send back
the resulting forecast to the U.S.
Thermonuclear fusion (核融合研究)
e.g., D + T → He + n + energy
臨界プラズマ条件: D-T反応(重水素と三重水素の反応)
では「発電炉内でプラズマ温度1億℃以上、密度100兆個
/cm3、1秒間以上閉じ込めることが条件」
核融合科学研究所 大型ヘリカル装置計画
http://www.lhd.nifs.ac.jp/
LHD(Large Helical Device、大型ヘリカル装置)は核融合研究の為
に日本の自然科学研究機構核融合科学研究所(NIFS)のLHDプロジェク
トによって製作された大型のヘリカル・ヘリオトロン型のプラズマ装
置。日本独自のヘリカル型磁場方式が用いられ、1時間以上にも亘る
長時間のプラズマ持続や、1016 個 / cm3 の高密度プラズマを成功させ
た。
AI applications
GA to determine optimal parameter set for magnetic field
configuration
NN to perform fine tuning for thermonuclear reactors
e.g.,
magnetic field configuration
intensity and timing of application of electric and magnetic fields
density distribution of “fuel” (D and T)