Agent-based traffic simulation model (Mesa + UXsim)

[EwoutH]

I'm developing a hybrid Agent-based traffic simulation model, using Mesa en UXsim. It's based on a real-world road network extracted and processed with OSMnx. It's in very early stages, but development can be followed here:

• https://github.com/EwoutH/urban-self-driving-effects/tree/main/model

The initialisation of the UXsim world is in traffic.py, and adding demand, retrieving travel times and executing the simulation will happen in model.py (mainly in the step function).

First two TODOs:

• Add demand from agent mode choices (to regions)
• Get travel times back between areas to update expected car travel time

Once it's in further stages I will describe it more extensively, and probably open a PR to add it as an example. But for now a quick share for anyone interested to follow.

CC @toruseo

Early conceptual model

[toruseo]

Thanks for sharing, this is an ambitious project!

Get travel times back between areas to update expected car travel time

If you need travel time only (not route), you may use World.ROUTECHOICE.dist[i,j] that store the current instantaneous travel time from node i to j.

This may be out of the scope, but it would be possible to formulate this as a day-to-day model, in which the "Travel time & price lookup" is came from previous day's experience so that it does not require significant external variables.

[EwoutH]

Thanks!

If you need travel time only (not route), you may use World.ROUTECHOICE.dist[i,j] that store the current instantaneous travel time from node i to j.

This would be useful as a starting point! Maybe we can keep a second counter like historical_dist which keeps the latest (few) travel time(s) for each node combination of journeys actually completed.

would be possible to formulate this as a day-to-day model, in which the "Travel time & price lookup" is came from previous day's experience so that it does not require significant external variables.

If both computational time and development time allow it, this is exactly what we plan on doing.

[EwoutH]

Quick update: Model is running and has reached minimal-viable product status, with a latest interaction bug between Mesa and UXsim resolved (correct bookkeeping of time in both is important): EwoutH/urban-self-driving-effects@2a702c3

The modal is simulating traffic in a ~half a million city during a full day, with up to 50 thousand vehicles simultaneously in UXsim. Each 900 seconds the model recalculates all travel times between all nodes on which allows Agents to take new mode choice decisions.

First traffic patterns are following expectations. Lot's of "magic numbers" of course, tomorrow is data collection, aggregation, plotting and validation day. Goal is to determine where to refine more.

Then it's taking a step back to the original research goals to see which questions we can exactly answer, what we need to be able to do that, which are unfeasible.