faq 281411594

Modal split in first iterations

by Johannes Lieberherr on 2018-05-23 15:19:28

---

Hi,

we want to analyze the demand effect of small/local changes of supply with MATSim.

For this we have a well converged MATSim reference run, from which we use the plans as initial state for our variant run (which contains the small/local changes). We call this "warmstart"-run.

Unexpectedly, the modal split changes very much in the first iterations, see the attached modestats-file. We reach a stable state only after 200 iterations, which seems to much for a "warmstart"-run with only small changes in supply. Since we start with a well calibrated state we would rather expect the modal split to change only slightly in the first iterations and converge fast.

Does anyone have any idea what might be causing this or what we might be doing wrong?

---

Comments: 3

---

Re: Modal split in first iterations

by Ihab Kaddoura on 2018-05-24 09:59:41

Hi,

Maybe you are using the selected plans as input instead of the output plans which contain all plans per agent (selected + non-selected)?

Another idea is, that the previous (well calibrated) run is quite old and some default parameters have changed; or you have set some parameters differently. Maybe, just compare the output config files (previous run vs. warmstart run)...

---

Re: Modal split in first iterations

by Kai Nagel on 2018-08-06 06:33:34

I am actually seeing similar things. A possible interpretation would be the following:

• Because of the non-innovative iterations 900 to 1000 of the previous run and the higher car share there, the car scores refer to a somewhat congested car system. (*)

• When you re-start, you have innovation switched on again, which implies that the "jump" from iteration 900 of the previous run is reverted instantaneously. The reason roughly is that some share of your population will, by the mode choice innovation, be forcibly thrown into some other mode. So car will go from 0.52 to 0.47, etc. In general, as long as mode innovation is switched on, you always have the according fraction of agents travelling on a mode that they don't really want (situation up to iteration 900 in your plot).

• From then on, why it is overshooting, I cannot really say.

  • Maybe, as Ihab suggests, you have only kept the selected plans, in which case it will take some time until agents drift back to their preferred plans. (Otherwise, a "bad" plan generated by mode innovation will immediately be removed again in the following iteration.)

  • If it is not that, then it becomes speculation. It also depends on if you are using SelectBest or SelectExpBeta. It also depends on how you are modelling the ride mode (we have in our recent simulations removed the ride mode from innovation at all, i.e. trips starting with it will just always keep it, because all current options to model that come with disadvantages).

Maybe attach the logfile, maybe then I can say more. :-). But it will remain speculation; somebody should do a research project about the population dynamics of MATSim.

---

Re: Modal split in first iterations

by Patrick Manser on 2018-08-07 05:52:43

Hi all!

Thank you for the answers and your interesting thoughts. We had a closer look at this
issue and found out that it actually was the removal of non-selected plans. Handing
over the full plan-set from the previous simulation run to the agents, the
height of the jump is more or less equal to the reversed jump at the point of
the innovation switch-off. After the jump in the first iteration, the curve
followed a constant line for all the modes (which is what we expected).

However, the question remains whether it is desired to let the agents have a full plan
set at the beginning of a simulation. A full plan set might contain misleading
information in the case of major modifications in the scenario setup, but the
mode statistic curves show a much more stable progress. On the other hand,
reducing the plan set to the selected plan only results in “wasting” 200-300
iterations for filling up the plan set with reasonable plans. But the agents
might better optimize their plan set to the new situation. Still, we don’t have
a satisfying answer for this question. But our experiments with different
initial plan sets show that the results might vary substantially.

Also, I like the idea of not allowing innovation in the case of the mode ride. We faced
some problems as well and struggled a lot calibrating it. On the other hand,
the ride mode is very sensitive to changes in the pt supply (for agents not
having a car available, it is the only alternative for longer distance trips).
As a result, the ride mode must be well calibrated in the plan-generation (before
the MATSim run starts).

Thanks again! And if anyone wants to share their experiences with the innovation
switch off process or with different initial plan sets (full plan set vs.
selected plans only and also vs. uncalibrated initial states), you are more
than welcome!

Best, Patrick (SBB)