Hi All
I'm trying to implement the neural oscillator described in this article
http://www.ecila.org/ecila_files/content/papers/ACEICMC05.pdf

It's basically 5 equations describing a pair of coupled neurons (or rather a simplified model of two coupled neurons). The author of the paper implemented in C++ , but I want to implement as a PD abstraction since my C++ is pretty well nonexistent and for the moment at least I just want to get a feel for what the possibilities are. I tried to implement in MAX/MSP 4, but there is no expr~ external for the windows version and I got nowhere trying to do the calculation with the maths objects. I have an excel spreadsheet doing the calcs and the output behaves the way I'd expect, so I'm fairly certain I've got the calcs right, It's just implementing in PD that I'm having trouble with.

SO...
The equations (bottom of page 1) are iterative - each new set of values for the equations depends on the output values for the previous step - eg X Y and Z at time n are calculated on the basis of the values X Y and Z at time n-1.
This leads me to the first question - since this means I need to do stepwise calculation, presumably I'll need to use fexpr~ (sample wise calculation) rather than expr~ (vector calculation)?

Or have I misunderstood? expr~ does vector calculations - and I'm assuming that's a displacement/amplitude vector - displacement from 0 over time t. If so, and since the time base is going to be constant surely I can just treat t as 1 - ie, ignore it - for present purposes?

That may be incredibly naive though - I'm new to expr~.

The second question is - does anyone have any ideas how I can start the thing running? Since it always needs a n-1 value to calculate the n value, I need to initialise it with values. Working on the principle that this is more or less a kind of waveshaping feedback synthesis I thought about just firing a short blast of white noise at all of the inputs to get the thing started.

Again, naive?

Either way, whether or not you have ideas to help - that paper is worth a look. Interesting stuff.

Ta
Dan

Hi

I did once a genetic algorithm with pd, using the pyext external. It takes a python class as an argument, and was well suited to boring-to-code-in-pd calculations. However, i had to smooth the output values as it was not running at audio rate.

Hi,

Did you have a look to how to solve Lorenz equation with [fexpr~] ? It's a subpatch located in the help file of expr. The process looks a lot like what you're trying to do.

ah - ta I'll take a look. RTFM I guess

I actually got it making a sound using expr~ the other night. I'm not sure it's "correct" as an implementation of those equations (I need to check a few things), but whether or not it is, it still makes some interesting sounds
I'll post it up when I get home.

I realised after I posted here that if it is neccessary to use fexpr~ then it might not be possible at all - because I need to calculate 4 different variables at each sample/step - so I'd need 4 fexpr~s and the ability to send sample-accurate data between them. I guess whether or not it's possible depends on whether it's sufficiently accurate to use [delay~ 1] to send the previous calculation step from one fexpr~'s out to another fexpr~'s input (this is exactly what I've done on my expr~ implementation).

Though having said that I just saw this -
"Expr, expr~, and fexpr~ now support multiple expressions separated by semicolons which results in multiple outlets"

Which might give me the answer...

And reading expr~ docs more carefully I realised that the vector calculation is based on the buffer size set in PD. So if I'm using expr~ I guess I should just set the delay~ to the buffer size instead of to 1 sample and calculate as vectors. Easier to work in fexpr~ though, clearly.

This all raises another question - does anyone know how the math objects in PD work? Do they calculate on a sample-by-sample or vector-by-vector basis?