You guys want to really predict earthquakes? Use an A.N.N. (Artificial Neural Network)

I've been reading thread after thread about earthquake predictions, and about the the sun's activity causing earthquakes implying the following causal chain:

Sun ejects stuff -> stuff punches magnetic field -> distorted field affects magnetic core -> disturbed core moves the crust -> crust movement triggers a release of stored energy in the form of vibrations

There may be other causal chains involved. These chains together probably form a causal web of complex relationships. We need to model that using data from a number of disciplines: meteorology, geology, whatever... there's tons of data, we can't wrap our mind around it. But as human beings, we tend to want to create models of these things. Why? To predict the future!

I know ATS is good at monitoring real time data from different sources from different disciplines. But how to model tons of data?

How about training an Artificial Neural Network to "learn" these relationships. It can't explicitly tell us what these causal relationships are, but I think it can tell us when an earthquake is likely to be triggered. That's good enough for us.

Earthquake predictions then would require 2 groups of people, which I think ATS has a lot of to a certain extent:
1. Data gatherers/organizers - those who are vigilant in monitoring various websites for realtime and historical hard data coming from monitoring devices like seismographs, magnetometers, etc.
2. Modelers - those who are willing to work their way around designing and fine-tuning models using neural network simulators (lots of downloadable open source NN models out there)

I see this as a collaborative task between members of ATS.

Or one can take the daunting task for himself and deal with collecting and organizing tons of data and at the same time working on the NN model. Or just stick with crude personal predictions based on limited data and non-optimized reasoning.

Anyway, the output of the NN should be: Time, Geographic Location, and Magnitude
The input will be a long list... the modelers should be able to determine which data are relevant, or maybe use the NN itself to decide that.

In general, we train the NN using historical data so it implicitly "learns" the relationships and forms a model based on that data (in the form of connection strengths between the nodes that make up the NN).

Once trained, throw at it present data and it will tell you a "prediction" based on what it "learned" from the past.

I'm sure it will work.

Collect data. Pick a NN model. Fine tune the NN in anticipation of the data. Train the NN using the data. Let is simmer. Input current data. NN will fill in the missing pieces ("predict").

I know nothing about geology, meteorology, neural networks, or whatever. So will a specialist from ATS please butt in and help us out?

Thanks.

Well, I have a bit of experience with a few different languages (mostly easy ones such as VB, Delphi, PHP, JavaScript), but my level of knowledge is no where near that required to work with NN frameworks. I would suggest though, that if this is done, many things should be fed into the equation. Earthquakes, volcanoes, all natural disasters basically, sun cycles, our position in the solar system, our position in the galaxy, biodiversity cycles, and so on...

PS - It's a great idea, I really like it. The problem will be finding people that are skilled enough. The data collection shouldn't be a problem.

Originally posted by WhizPhiz
Well, I have a bit of experience with a few different languages (mostly easy ones such as VB, Delphi, PHP, JavaScript), but my level of knowledge is no where near that required to work with NN frameworks.

Thanks appreciate the response. I don't actually think expert level is needed of everyone to participate. Maybe just some fooling around with a few downloadable open source NN apps will get us started. No need to program really, you can just play around with these "fun toys" (they're like mini brains) with some inputs, outputs, maybe play around with the number of nodes and layers, etc., and see what happens.

Of course some bit of background reading would be best, but whatever time permits. Either way it's at least fun to play with.

I would suggest though, that if this is done, many things should be fed into the equation. Earthquakes, volcanoes, all natural disasters basically, sun cycles, our position in the solar system, our position in the galaxy, biodiversity cycles, and so on...

Yes there's a long list of possible inputs. My guess is the most relevant ones are those with cycles that are close to our timescales. Cycles that span millions of years (like galactic orbit) would probably be less of a factor in determining events at our normal time resolution, I would guess.

PS - It's a great idea, I really like it. The problem will be finding people that are skilled enough. The data collection shouldn't be a problem.

Thanks. I suggest we just play with these fun toys and see what happens, until the specialists arrive.

This link might help get you started:

Neural Networks at your Fingertips

There's a lot of NN toys out there from other sites to play with using input and output data from our daily lives, or tweak the source code with. This can be a new hobby for ATS members, no kidding.

No need to wait for the experts I'd say. But it would be great if they join in, we need guidance.

Originally posted by laymanskeptic
Earthquake predictions then would require 2 groups of people, which I think ATS has a lot of to a certain extent:
1. Data gatherers/organizers - those who are vigilant in monitoring various websites for realtime and historical hard data coming from monitoring devices like seismographs, magnetometers, etc.
2. Modelers - those who are willing to work their way around designing and fine-tuning models using neural network simulators (lots of downloadable open source NN models out there)

edit on 3/30/2011 by laymanskeptic because: (no reason given)

edit on 3/30/2011 by laymanskeptic because: (no reason given)

So on second thought, in the absence of experts, I encourage you guys who have access to historical and recent data to play around with NN apps.

reply to post by laymanskeptic
S&F ..I think that would be a good idea ....We seem to have knowledge that humans act out around full moons and animals go crazy before earth quakes! having people putting data into the system would be a good hobby for some of us ...zoo page, bartender page compass and geomagnetic page,etc ....it could work...peace & thanks for making the suggestion .

I think this is going to be a long-term project for us ATSers. On and off.

No immediate results, no instant gratification. A lot of data gathering and organization. A lot of fooling around with the different NN models. Maybe after we get a feel for it, someone decides to write his own code designed for our purpose. Etc...

A lot of testing against real world events.

Originally posted by the2ofusr1
reply to post by laymanskeptic

 

having people putting data into the system would be a good hobby for some of us ...zoo page, bartender page compass and geomagnetic page,etc ....it could work...peace & thanks for making the suggestion .

Yeah you're right there could be a database that ATSers can access to input their data, a different page for each data category.

I hope some coder (or group of coders) here would take on the challenge.

Maybe it can be a web app on Google App Engine or whatever.

Oh wow, this thread crashed before it even took off... lolz

Makes you wonder...

EDIT: naaaah... to early to tell


EDIT: and here I am talking to myself

Have the Governments around the world not got this idea already running on one of those super computers they keep ?,i think its a really good idea sad to see the thread not get much attention but people would rather look at "my wifes a alien from Venus" titled threads than actually read a thread that might hurt their brain good luck and i hope the right people find the thread change the title lol