Non-Determinism + Random Spikes in Data Plots

The NRP does not compute simulations deterministically, running the same simulation two times different results are obtained. In addition, depending on the compute load of Gazebo and NEST different quality of resulting plots are generated, in particular random spikes may occur in the data.

Setup:
A sin(t) signal is generatd in the Transfer Function, fed into three identical LIF neurons in PYNN. The neural output is retrieved and assigned to a 1DOF revolute joint model in the NRP.

1) runnin the same simulation two times generates different plots of the resulting joint angle of the model.
See:

2) We add 1000 additional not connected neurons to PYNN to increase its workload. Afterwards without the additional neurons, we add a sleep(2) for two seconds in a Transfer Function:

Despite the exact same experiment setup, both these plots have a much better quality, in particular we do not see random spikes occuring.

The results may show the deficies of the CLE simulator calls, in particular if the procedure is implemented as
------ Gazebo ------
– NEST – – TF –
an increased workload in the brain pipe can lead to this:
------ Gazebo ------
— NEST — — TF —

Instead, simulator and Transfer funtion calls should always synchronize properly like:
------ Gazebo ------ — TF —
— NEST —
and
— Gazebo —
------ NEST ------ — TF —

Research experiments multiple simulation runs need to aim the exact same experiment result.

Benedikt