Schedulefeature

[auroraschmidt]

This is an addition that allows for PoissonScheduling (note that times in the poisson schedule are still rounded to integers). It is used in the simple_contagion model in the Tsunami repo

[coveralls]

Coverage increased (+1.8%) to 90.52% when pulling dc96b67 on schedulefeature into 7145a42 on master.

[cash]

I'm not sure why we are calling this a Poisson process. Isn't this a Bernoulli process now? And wouldn't it be more efficient to implement it that way with a geometric distribution (for small lambda, this throws away 10-40% of the samples)?

[auroraschmidt]

A Poisson process has exponential inter-arrival times. We are sampling inter-activity times with the exponential distribution, so that's why we are calling it Poisson.

[cash]

But these aren't exponential wait times. They are geometric wait times where

Here's a sample plot at lambda=5 and p=e^-y (discretized exponential and geometric):

[cash]

@auroraschmidt I've added two implementations of a Bernoulli scheduler. They both give the same arrival time distribution as the original PoissonScheduler implementation. BernoulliScheduler is slower than the original implementation but handles indefinite simulations. The FastBernoulliScheduler works in the same manner as PoissonScheduler by pre-generating the schedule.

Do you have strong opinions on whether the stop time for the scheduler should be inclusive or exclusive? Right now if you configure with a start time of 0 and stop time of 5, you get 4 time steps of 1, 2, 3, 4. I'd lean toward inclusive.