|
PureBytes Links
Trading Reference Links
|
Dans un courrier daté du 24/09/98 18:45:36 , vous avez écrit :
<<
Please, could someone define these terms so that we may know what the
differences are?
>>
Ok, I admit that I have used some shortcuts...
NN = Neural Network
NN by Genetic Algo = NN trained by a Genetic Algorithm
Backpercolation NN = Proprietary method to train these NN (Originator: Mark
Jurik)
NN algorithms = The structure of the neuron synapses, if you want ( dozens of
variants). How neurons are connected to the others if you prefer. Thisis the
structure of the artificial "brain" coded into computer language.
Kohonen Nets= NN that makes classification. Also known as LVQ (LinearVector
Quantization) or unsupervised learning.
recursive nets = NN where connections are made from the output(s) to the
intermediary inputs. Difficult to explain without a chart. The idea is to
reinject a part of the answer produced by some neuron into the inputs slab
that feds the neurons. Several structures also...
GRNN = General Regression NN (by Dr Donald Spech)
PRNN Nets = Probabilistic NN (by Dr Donald Spech)
NN evolved by GA = NN trained and build by use of a genetic algorithm.
multidimentional Nets, Proprietary NN where all neurons are or can be
connected (no slabs), like in your brain.
Scorecards, A calssification method that gives a ponderation to an input
according ti its membership to the cell it belongs to.
Notes are summed and compared to a threshold value.
This technique is widely used in credit permission (Now, you know why your
bank asked how old you are, if married , how many chikdren and so on. It to
fed a Scorecard and accept the credit or reject it if your global note is
below the threshold).
Rule bases, = Decision tree (if A then do else if B the do else ifC then do
, and so on , with multiple branchs (like a tree) ).
all evolved by very sophisticated GA = Build by a top notch genetic algorithm.
There are a lot of variants in this technique, and what we have used was the
best of the technology available at this time, far beyond the classical way
to use GA operators).
neurofuzzy logic = A mix of fuzzy logic and NN.
More in depth explanation on www.sirtrade.com
RMSE error = Root Mean Squared error. It's the prediction error obseerved
when running the predictor (a NN is a pedictor, a scorecard is an other) on
data.
evaluation function = Used with GA techniques . When evolving solutions , you
need to give them a value , accorfding to the goal to reach (could be net
profit for a trading system). Then this allow to discard some soutions, to
mate soe other preferentially , and so on.
We also use this with neurofuzzy logic.
You do not need to know all of this to use a modern AI software....
Sincerely,
Pierre Orphelin
|