How to Structure your Simulations

This project was born out of the need for a tool to manage and store results of spiking neural network simulations. In particular, to manage results obtained with the BRIAN neural network simulator. The more complex simulations become the more complicated are the sets of parameters and the harder it gets to qualitatively compare results obtained from different simulation runs. There was a huge need to standardize the simulations and parameter exploration, to structure and group different parameters, and especially to entangle the scientific simulation from the environment that runs it!

I would like to present you some ideas on how to effectively use this tool. I would divide any numerical simulations into 3 phases:

1. Parameter Definition Phase

   Here you will only add parameters (see pypet.naturalnaming.ParameterNode.f_add_parameter(), and Addition of Groups and Leaves (aka Results and Parameters)) to your trajectory, no results, no derived parameters, no building of other objects used during your simulation. ONLY parameters. You could write a conf.py file that adds all parameters to your trajectory. To exclude parameter sets and to allow some control flow, you can consider Presetting of Parameters.

2. Preparation Phase

   Here you will prepare stuff before the actual runtime, e.g. create objects needed in your simulations. This encompasses stuff that only needs to be build once and is used during all individual runs. Here you can also start adding derived parameters.

   At the end of your preparation phase you define which parameters should be explored and how via the pypet.trajectory.f_explore() (and take a look at Parameter Exploration).

3. The Run Phase

   This is the phase where individual parameter space points along the trajectory that you explore are evaluated. Here you produce your main numerical results and maybe some derived parameters. You have a top-level function that uses a single run object (see Single Runs) and accesses the parameters needed during the single run to make some calculations.

   This top level function is handed over to the runtime environment in addition with some other arguments (like some objects not managed by your trajectory) to carry out the simulation (see pypet.environment.Environment.f_run, and :ref:`more-on-running()).

   Usually to speed up your simulations and to compute several runs in parallel, you can use multiprocessing at this stage, see Multiprocessing.

After your parameter exploration is finished you might have a 4th stage of post processing. For instance, calculating summary statistics over all your simulation runs. Yet, I would separate this phase entirely from the previous ones. You can do this in a separate program that loads the trajectory.

Well, that’s it, so thanks for using pypet,

Robert