Re-learning Matlab
As well as defining a project for me to do, I have to pick a platform to use for the actual experiments. I'm naturally drawn towards write stuff mostly from scratch in C++, using a few header files people have supplied
, because that's how I'm most used to working, but it may not be the right approach here. When I was working with CTRNNs it was natural, because Randy had supplied me with headers that implemented a CTRNN very efficiently, and provided a very flexible basic structure for the evolutionary algorithm. Now that I'm likely to be working with a discrete type of network model, the closest such resource I know of is Ilya's BN/PBN Matlab toolbox, so I'm giving Matlab a try as a platform.
I have used Matlab before—a component of my MSc involved using it to program neural networks—but haven't actually touched in about 5 years simply because my lab mostly uses Mathematica and it's been convenient to use the same tools as my advisor. So I find myself having to do some re-learning, though fortunately it's a sensibly designed language and not too hard to get back into. It also helps that the documentation is very good and has some nice tutorials to get started with; I had always had this nagging feeling that it took me less time to get to grips with Matlab than Mathematica, and now I realise that this is why.
It took me a day of messing around to remember roughly how to interact with the Matlab environment, and I'm spending today getting to grips with the BN/PBN toolbox. I think I've reached the point where doing more toy exercises won't help very much, so it's time I set myself a little task to apply this. I think I'm going to try and replicate the yeast cell regulatory network from Li et. al's paper, and then try the same thing in C++ to see how the two environments compare.
While I'm on the subject, I should note a few other useful downloads. The visualisation functions in the BN/PBN toolbox depend on a graph layout package that didn't seem to be working on my install, but A. Taylan Cemgil's graph layout package did the trick. Also, many people complain about the slowness of Matlab for simulations with many iterations, so I should probably see if Tom Minka's lightspeed toolbox will run on a non-Windows system (it sounds like it should, but he hasn't tested this).
Comments