outrun wrote:(...) An energy company I worked for 15 years ago used SVM and kernel methods.

(...) but non of the asset managers I know.. some did experiment with it though.

That's interesting.

Can you give some hints on the kind of predictions based machine learning techniques at that energy company?

Why that non-adoption of ML techniques by asset managers?

We used support vector regression to predict electricity demand (of our customers) on a 15 minute bases, starting right now, till 7 days ahead. We had various explanatory inputs like time of day, day of week, upcoming crazy days, light intentisty predictions, temperature, windspeed,..

We did the same for electricty spot prices, the demand predictions were inputs, we also modellen the demand of the country, and all renewable production forecast. We also had Windparks and had to predict the production of those. And then some more, like high resolution electricty and gas forward curves (hourly mean prices for the next 5 years)

In those days kernel methods and support vector machines were state of the art. Nowadays the first pick for this task (imo) would be random forests, or maybe neural networks if you want to model the dynamics of the time series.

Statistics: Posted by outrun — Today, 12:21 am

]]>

(...) An energy company I worked for 15 years ago used SVM and kernel methods.

(...) but non of the asset managers I know.. some did experiment with it though.

That's interesting.

Can you give some hints on the kind of predictions based machine learning techniques at that energy company?

Why that non-adoption of ML techniques by asset managers?

Statistics: Posted by tagoma — Yesterday, 5:47 pm

]]>

]]>

]]>

Then there's things like Intel's USB NN stick (http://www.anandtech.com/show/11649/int ... pute-stick ) which seems to be targeted at product developers that might want to imbue a little deep learning into a product with a special chip ( https://www.movidius.com/solutions/mach ... e-learning). That might be fantastic for a bunch of nearly-invisible and almost ubiquitous applications in home appliances, automotive, toys, industrial control, facilities management, etc.

Too often, technologists and pundits fixate on whether some invention is going solve the "big" problems of the world -- cure cancer, end world hunger, etc. They conclude it can't do it and that the technology must just be hype. Meanwhile, the actual power of the invention comes from addressing a large number of inconsequential applications -- a better cup of coffee, a vending machine that anticipates demand, a washing machine that automagically removes stains, a laptop that gets another hour of battery life through smart power management.

Statistics: Posted by Traden4Alpha — Yesterday, 4:16 pm

]]>

Whether "deep learning" is being used in business or sought by employers is an empirical question. An analysis of resumes on LinkedIn and job postings in various internet sites would provide an answer.

As for the efficiency argument, he's technically right but realistically wrong on two levels. First, inefficiency is irrelevant in an environment of extremely cheap high-performance computing. All of today's major OSes are horribly inefficient relative to the simpler predecessors but that's not slowed their adoption. Today's mobile device OSes require a 1990s supercomputer to run and they have it. I'd bet every major corporation has a ideo card toaster somewhere and if they don't they can rent one from a cloud vendor.

Second, finding people who have his laundry list of techniques (and actually know how to use them) is probably a lot harder than finding someone who knows how to spin-up an instance of TensorFlow and throw a bunch of data at it. It feels like the world is evolving from one where engineers understand every line of code in the system to one where they just plug in a few extremely powerful black boxes that just work (most of the time).

Taking these (and possibly other) points into consideration, where is this technology as we speak? My gut feeling says 2.5% based on the assumption that production systems are still in alpha/beta stages.

Statistics: Posted by Cuchulainn — Yesterday, 2:36 pm

]]>

As for the efficiency argument, he's technically right but realistically wrong on two levels. First, inefficiency is irrelevant in an environment of extremely cheap high-performance computing. All of today's major OSes are horribly inefficient relative to the simpler predecessors but that's not slowed their adoption. Today's mobile device OSes require a 1990s supercomputer to run and they have it. I'd bet every major corporation has a ideo card toaster somewhere and if they don't they can rent one from a cloud vendor.

Second, finding people who have his laundry list of techniques (and actually know how to use them) is probably a lot harder than finding someone who knows how to spin-up an instance of TensorFlow and throw a bunch of data at it. It feels like the world is evolving from one where engineers understand every line of code in the system to one where they just plug in a few extremely powerful black boxes that just work (most of the time).

Statistics: Posted by Traden4Alpha — Yesterday, 1:01 pm

]]>

Statistics: Posted by Cuchulainn — Yesterday, 8:44 am

]]>

. The West and Quantlib implementations of the Genz algorithms give the same accuracy up to machine precision. We used the finite difference method to check them just to ensure that they do not give the same incorrect value for certain values of the parameters. We generated a,b,rho randomly and computed 10^10 trials. Use FDM to check both Genz algorithms.

. The added value of the finite difference approach is that it can be tuned to suit our accuracy needs. The basic model is second-order accurate and we can use Richardson extrapolation to achieve fourth-order accuracy. Furthermore, these meshes deliver a matrix as part of the algorithm which saves recomputing for certain classes of applications.

. The finite difference approach can be applied to a wide range of bivariate and trivariate distributions, for example the bivariate t distribution and the trivariate normal distribution. Genz becomes difficult in these cases.

. The Drezner1978 algorithm is less accurate than the other schemes and it degrades for correlation near 1 in absolute.

Statistics: Posted by Cuchulainn — Yesterday, 8:25 am

]]>

In general a NN is a flexible parametric function, and the precision is estimated empirically instead of analytically.

Would be nice to investigate efficiency: nr of parameters vs precision. NN are not very efficient, they have *many* local minima, but people like that! It acts as a regularization factor that prevents overfitting. If you *do* want more accuracy then you increase the network size, not the search for the optimum. There are also explicit regulization methods (like penalties for high L1 or L2 sums of all the parameters) that change the optimisation surface shape and can remove local minima. It's all very enpirical, but in a explicit way. Using eg splines, polynomial is often motivated by the properties of those method (eg smoothness), but I don't see many people actually check if those properties are in the data. Or when doing a chebyshev approximation you'll often know worst case error bound of that approximation which is really nice, but that's for approximating well known analytical functions, not sets of data points (you'll never know if new data will mess up you theory about the current data you're fitting on). In that sense exp5 is a bit unconventional for NN. It's not a "data driven" problem, one would eg need to first generate many {x,exp(x)} sample pairs and use those to train the model to fit exp(x) on some range of x values. Very much like regression really..

Statistics: Posted by outrun — July 19th, 2017, 9:29 pm

]]>

DE and GA do this very well.

Statistics: Posted by Cuchulainn — July 19th, 2017, 8:28 pm

]]>

https://en.wikiversity.org/wiki/Introdu ... _functions

My hunch is with ANN you have to retrain for all of Antarctica? Yuge. Can you teach NN to speaka the cubic splines without rummaging through bitmaps?

Follow-on Q: why and when to train? i.e. under which circumstances.

// Vax 11/750/780 and Vax/VMS best OS ever.

Statistics: Posted by Cuchulainn — July 15th, 2017, 8:12 pm

]]>

]]>

]]>

I have seen image matching systems that did use splines. It was a project at a former employer that was for tracking sea ice in the Arctic for JPL/NASA. The challenge was to recognize and track the icebergs and floating plates of ice across months and years as they wandered around the ocean & icepack. The advantage of representing the outline of each major ice floe as a spline was that was that splines more robust to changes in ice object shape. Unlike many curve-fitting methods in which each coefficient of the curve fit is a function of ALL of the data, spline coefficients are only affected by a local neighborhood. If an ice feature lost a chunk, the spline coefficients in the remaining part did not change and could still be matched. This was back in the mid 80s, ran on a Vax 11/750, and was definitely an example of heavy thought by human engineers rather than throwing terabytes of data onto some automagical adaptive learning system on a GPU array.

Statistics: Posted by Traden4Alpha — July 15th, 2017, 4:32 pm

]]>