|
I have a state described by 3 continuous numerical variables a, b and c that vary continuously. At some point, the combination state {a, b, c} will cause the emission of a binary signal, and then a,b, and c get reset, and the process repeats itself. I want to learn(offline) a function f(a, b, c) that soon enough before the signal emission, will help me predict it. One possibility is to have continuously updated probability distribution over the 2 emitted signals. Other possibility is to have threshold-type rule that says for example f(a,b, c) > X makes one prediction, and f(a,b,c) < -X makes another prediction, and in between we don't know(no decision region) What is a simple, practical, applciable framework I can work in to approach this problem? Any recommended techniques? P.S. The state {a, b, c} varies continuously and is reset each time a signal is emitted. Signals are emitted at arbitrary discrete times unknown beforehand. |
|
Out of the top of my head, you might generate some training data, where you will have:
Right there you then have a classification problem, for what you are describing, it is also a pretty unbalanced classification problem, but with only 3 features (a,b,c) you should be able to use logistic regression or SVMs |
