4. Tuning window, the number of time-step, for behavior classifier (or pattern recognition)

[hyunjimoon]

Yaman's team named the window length as segment and set this as 12 from the code (ISTS)).

Pseudo-code of the classifier by Kanar (1999)
Receive the dynamic behavior, X, to be classified
Normalize X, and obtain Xnorm
Split Xnorm into equal-length segments, and obtain S,, Sz, ... S,
Iterate over all segments {
Pick a segment S;
Calculate mean for the segment, i.e., a;
Calculate slope for the segment, i.e., b;
Calculate curvature for the segment, i.e., c
Iterate over all defined behavior classes (
Pick a behavior class, B
Calculate the state-optimized likelihood of norm belonging to the B, class,
i.e., Pi
}
Calculate the maximum likelihood, Pmax=max(p)
Report the behavior class j for which the p=P max

[tomfid]

I think the model time horizon definitely influences at least our perception of behavior, or our ability to discern what is going on.

Sin wave or logistic growth? What if there's measurement error?

I think a closely related challenge is the selection of a window for behavior classification. One could, for example, further decompose the trajectories above into 2 phases dominated by + and - feedback.

I'm guessing these questions are addressed in the BATS papers, but it's been a long time since I read them.

[hyunjimoon]

Similar spirit: Ridge or Multiple peak

I am not saying ridge and multiple local peaks are the same. However, our inspection can be the same. e.g. you can get the exact same samples (3,4,5) from two different distributions: gaussian mixture (.5 Normal(3, 1) + .5 Normal(5,1)) and Uniform (3,5). First has two local peaks, the second is a ridge.

Highly correlated parameters (e.g. your example) are not the same as multiple local peaks, though when you have them you get a ridge and you should essentially remove one parameter from optimization (i.e. fix/assume it). Local peaks are best identified by looking at convergence of multiple restarts in optimization to different locations/peaks.