search

piLaboratory / GillesCom

Gillespie Simulation of Ecological Communities Dynamics
2 stars 0 forks source link

sads convergence? #6

Open andrechalom opened 8 years ago

andrechalom commented 8 years ago

We need to study how can we decide if the sads has converged, in order to implement a function of "run bdm until convergence".

andrechalom commented 8 years ago

Thinking about convergence, we can try to implement a measure of divergence between subsequent "snapshots", and test whether (and at what rate) they are converging to zero. Some questions that must be answered:

  1. What is the best choice of divergence? Statistical distance? Kullback-Leibler? Jensen–Shannon (which is symetric)? Some rank transformed distance (probably akin to sbma, as we don't really care about the ordering on the tail)? Or some parametric estimate related to sads, such as Fisher's alpha difference?
  2. How does the snapshot frequency change the perception of convergence? Taking too much snapshots might gives us the false idea that the sads has converged? How can we avoid this trap?
andrechalom commented 8 years ago

3- The amplitude of variation for the statistics seem to be fairly large, even when the community composition is fairly "stable". How much noise are we willing to accept into the "convergence"?

Not exactly related to this topic, but worthy of mention: the default values for a community of 100 species seem to produce a community that follows a weibull distribution (or a bs/geom), with fitls producing an AIC 100 units larger than these.