faq 29720608

How stochastic are travel times in MATSim?

by Kai Nagel on 2015-06-18 17:19:15

---

Proxying for Michal Maciejewski:

One postdoc from DTU (Denmark) was asking about the stochasticity -- he wants to directly include stochastic definition of travel times into his optimization problem. Should he run different days to get different values, or maybe 1 day with different random seeds?

---

Comments: 1

---

Re: How stochastic are travel times in MATSim?

by Kai Nagel on 2015-06-18 17:25:43

Nothing systematic available, unfortunately.  Two approaches seem to make sense:

(1) run same day/iteration with different random seeds.  Effect will _only_ be the stochasticity of the mobsim, which currently is fairly small (flow is accumulated deterministically, but priority at intersections is (still?) stochastic)

(2) run different days/iterations.  I would say that this is best done with innovation switched off, and with  SelectExpBetaPlanStrategy (and not ChangeXxx, which we normally recommend; "Select" makes a new full logit draw in every iterations; Change... only changes _some_ of the plans.).

Also the scoring should be switched off.  Probably easiest achieved by switching the learning rate to zero.

Both things should be tested; if done right, the plans should not change over the iterations except for the "selected" attributes.

This should have more stochasticity than (1), based on variations in plans choice between days/draws.  It might look like this has more variation than using ChangeXxx, but in the end it should not, the only difference is correlations: ChangeXxx is correlated from one iteration to the next, SelectXxx is not.