The deadend trimming is too enthusiastic, getting rid of some unsnapped
cycleways and things connecting to the map border. Will iterate on it
this week; net benefit for now.
intersection. This often happens with a group of 4 intersections (two
divided highways), and there may be many small segments embedded in the
middle for street car tracks and such.
Also bring in fresh OSM for Tempe, with one such intersection now
consolidated! #654, #672
- Grab fresh Seattle OSM, picking up https://www.openstreetmap.org/changeset/108071529
- Treat highway=footway, bicycle=dismount as a cyclepath, for now
- Treat service=driveway, bicycle=designated as a cyclepath
Since this requires regenerating all maps anyway, also include some
stuff to improve Aurora near Green Lake:
- stop making highway lanes super wide by default; they just make
divided highways overlap themselves
- filter out service roads with access=customers
But note the bridge from the Arboretum to Lynn is still disconnected,
because of detailed footway mapping that isn't tagged as
bike-accessible.
dramatically improve time to import and edit maps.
The fix helps all maps that use extremely high edge weights to prevent
people from cutting through private roads. There may be a more robust
fast_paths fix later, but I want to reap the benefits for tomorrow's
release.
The dramatic numbers:
- importing huge_seattle: 893s down to 108s
- editing huge_seattle: 102s down to 19s
Query speeds didn't appear to substantially change.
lane that they're stuck behind them. Only record a risk exposure event
the first time, but let passing happen anywhere. #382
Also add scenario name to PrebakeSummary, to disambiguate the Poundbury
results.
fixes a very dramatic problem in the Green Lake map.
Regenerating everything...
Also added total trip time to the prebaked summary, to get a quick sense
if a change net helps or hurts and have a record in version control.
lane and both want to park in the same building or parking lot, they're
assigned the same spot. When the second one tries to use it, they pick
another. Now the second will still consider spots that they're directly
next to. #688
- handle when the equiv_pos of a driveway gets too close to the edge of
another lane
- make the updater workflow handle files from S3 that're a bit older
- remove pathfinding_avoiding_roads
- strip out old vehicle capping from map edit JSON, then fix up
proposals
- delete old capping API example
- temporarily give up on phinney; it starts gridlocking
- add broadmoor proposal link in-game
supported!
- Had to skip over center turn lanes -- we're approaching the point where
we can model those realistically.
- Carefully deal with static blockages near the start of a lane to avoid
spillover. Observed near b3810 in greenlake. It's time to make equiv_pos
smarter...
- offstreet_parking_length in importer config
- backwards compatibility for map edits
- fixing up the baked-in proposals
- working around a few PolyLine bugs that happen with the new rounding
- Don't regenerate actdev scenarios yet -- the upstream JSON format is
out of date, will have to fix separately.
I have some fuzzy vision of declaring all importer pieces as a dependency graph, with optional caching of some input, that would express this kind of thing better. But... in the meantime.
Now regenerating everything...
- Allow blocking the box around two complex intersections in Green Lake.
This makes the vehicle behavior much more realistic there, by visual
inspection.
- Amp up offstreet parking to 10 per building. I noticed the simulation
completes easily with --infinite_parking. This is an approximation of
that. We make really bad guesses about carpooling and the amount of
parking available around here, so effectively just remove it from
consideration for now.
- arrays are now iterable directly
- switch to using BTree{Set,Map}::retain!
- a round of clippy
- regenerate scenarios and prebaked data; not sure why, but there's a
diff
- Support this at the pathfinding level, when transforming v2->v1
- Adjust how the vehicle's body is rendered as it exits a driveway onto
a farther lane
No support yet for blocking any intermediate lanes; vehicles may clip
through each other without any conflict. Planning to add that
separately.
Regenerating all scenarios and prebaked data...
... But leave it disabled, until we can handle such larger file sizes
and levels of traffic.
And although this should be a behavioral no-op, there's a diff in the
scenario files, so recalculate them.
- Stop importing rail in Tempe. Not simulating anything on it yet, and it complicates gridlock. #672
- Update the GMNS timing.csv import based on slight format change. #626
running out on my current machine. Fixes#671.
Finally regenerate screenshots for the first time in ages... just
blindly accepting everything, because the slightly different screen size
means everything was slightly shifted down.
count incoming roads when figuring out if an intersection is degenerate.
Also make link roads (on/off ramps) lower priority than the main part of
the road.
Regenerated everything.
(and fixing up the cloud scripts)
times in the past, I've also tried doing this for other roads, but wound
up reverting for reasons only git remembers. But it seems like an
obvious win for bike paths; especially around Seattle, the ways are
split because of all of this raised curb data we're ignoring anyway.
Tested manually around Montlake.
Finally regenerating all data... Only Phinney breaks. One for tomorrow.
https://www.openstreetmap.org/changeset/105381427 mapping a turn lane is
the fix, preventing crazy U-turns using the service road. Not sure how I
missed that lane when looking here before. I made the edit to .osm
locally instead of grabbing fresh data for all of Seattle.
for people that leave one border, then re-enter a different one. #664
Alternative considered: insert a dummy remote trip between the two
borders. We used to do something like this at non-trivial code
complexity expense and having to explain the trip in the UI.
Regenerated all scenarios and prebaked data.
- Modest file size increase, as expected. Montlake scenario goes from
1.3MB to 1.5, downtown from 37MB to 43MB, all Seattle scenarios from
280MB to 327MB
- Eyeballed a few of the previously broken trips; they work now!
- Montlake goes from 3127 cancelled trips to just 302!
- No new gridlock, except in Rainier Valley -- disabling that for now
- I discovered missing validation in Poundbury for no-op trips between
the same building. I'll filter those out and restore prebaked results
there in a followup.
way, we have trip stats for people starting near the end of the day, and
we stop incorrectly showing failed trips when comparing.
Prebaked data: no change in size (245MB)
Montlake: 3135 "cancelled" trips -> 3127
Lakeslice: 6766 "cancelled" trips -> 6647
First time doing this in about a week. Trevor's new amenities should
show up, and lane widths are now adjusted. Haven't yet regenerated
the screenshot diff, because it's segfaulting (!!) for some reason. And
Seattle's popdat.bin and ALL of the scenarios are temporarily hosed. So
if you pull from git head and run the updater, expect some weirdness.
Working on fixing this.
No behavioral change here; this is a trivial transformation. If a
directed road has any walkable lane, then there's exactly 1 of them. I
verified by manually checking paths and also seeing prebaked results
having zero diff.
always tiny; Dijkstra's is fine. It costs a bit of file size to store
it. The huge leeds map goes from 160MB to 157MB -- not crazy savings,
but something.
Also fix a slight bug with 92d3a890ea that
caused some pedestrians to uselessly visit a bus stop node while
routing. (southbank crashes a few hours in otherwise)
This is simpler to reason about, allows the penalty for entering a zone
or taking an unprotected turn to be expressed in terms of a time
penalty, and is a step towards adjusting bike/foot routing for elevation
data.
When we later add things like "safety/quietness" for cycling, maybe we
can switch to using a (time, quietness) tuple, and transform into a
single number with a linear combination parameterized by that agent's
preference for time/safety. This change is compatible with that future
idea.
There are behavior changes here, particularly for zones and unprotected
turns. No new maps start gridlocking, and in fact, Rainier starts
working again.
"rise / run" calculation used the trimmed road center-lines, which don't
match up with the elevation at each original intersection point.
Also handle infinity in the output and reduce the resolution of the
query from every 1m to every 5m.
Regenerate all maps due to the map format change. Try bringing in
elevation data for all of Seattle using the LIDAR source, since
the data quality assessed in eldang/elevation_lookups#12 seems to be
similar, and LIDAR is way faster than contours.
Regenerate all maps. Gridlock-wise, Rainier and Poundbury broke, but
Wallingford started working again. Acceptable cost for a change this
useful; I'll work on fixing those maps later.
stop importing golf cart paths, even though they would be kind of
interesting to use for this proposal...
Interventions needed to keep lakeslice running, of course
Instead of just picking the intersectin closest to the origin or
destination, calculate the full path length, and take the one with the
shortest distance. This fixes some of the weird problems routing around
Broadmoor. Regenerate all prebaked data.
Also fix the original request for paths involving zones, so tracing it
later works.
So far, just call it for Seattle maps. Store the data sources in S3.
Note this'll only run on my machine right now, unless you also build the
Docker image locally. Failures in elevation should be skipped for now.
Before: time starts when the vehicle reaches the front of the queue and
first requests their turn
After: time starts when the vehicle first becomes blocked on the queue
leading to the intersection.
Regenerate prebaked data.
wind up looping back on themselves in a nonsensical way, causing
vehicles to visually glitch when moving through.
This was started in 081819d86b, but it
used to gridlock 2 maps. All the recent roundabout fixes seems to have
resolved those! And adjusting offstreet parking for two maps.
But wallingford does regress; plunging forward for now.
scenarios, so we can run A/B tests with map edits. cyipt/actdev#114
To fix it up, I hand-timed
https://www.openstreetmap.org/node/2124133019, which could have smarter
heuristics as a button-operated half-signal in the future. And allowed
blocking-the-box on some small intersections near that area.
1) If a car is blocked by a conflicting turn and is part of a cycle,
wake up the car blocking it. In some cases, this wakes it up faster
and unsticks things. Otherwise, it just wastes a little bit of time.
2) If a car is part of a cycle, allow blocking-the-box.
3) Continue sorting people at a stop sign by the time they've been
waiting. But for cars "overflowing" their current lane, move them to
the front of this ordering. It unsticks one particular situation.
4) Fix wakeup_waiting entirely. Before, it was waking up protected turns
before permitted, but otherwise the ordering was arbitrary. Now actually
respect stop sign ordering. I expect this to improve many other
situations than the one I was checking.
This was all motivated by one particular roundabout in Poundbury. It
doesn't solve gridlock there, but it gets past a major blockage.
If a study area exists for the map, make a copy of the base/active
scenarios with the background traffic mixed in. Also remove people
living in the site, since they're redundant.
Ran it like this: for city in `ls data/system/gb/`; do ./import.sh
--scenario --city=gb/$city || break; done
attention to which intersection is being destroyed. Fixes#527 --
montlake and phinney both look correct now.
Regenerating everything. Actually, Phinney now runs, so adding a 4th
prebaked map!!! But Rainier regressed -- there's an issue with the
signal heuristics that's now a problem; I'll fix later.
I experimented on the Rainier Valley map, which recently started
gridlocking due to too many cars doing this, to tune the value. Got it
running again! The two other maps keep running, with some trips on
average getting a little slower.
Added an extra step to classify service roads as running through a
parking lot, to prevent them from being treated as regular roads.
Had to fix up a few prebaked traffic signals. lakeslice falls back into
gridlock; will fix separately -- too much effort behind this change to
stop.
be more careful with nodes representing uber-turns. Even if that vehicle
type doesn't use an uber-turn, we still need to force the nodes to exist
and match up between input graphs.
Although this really only fixes gb/charleville_mezieres/secteur4, it
potentially affects all maps, because the node map changes. So
regenerate everything...
City names are now disambiguated by a two-letter country code. This
commit handles almost everything needed to make this transition. Main
next steps are fixing up map edits automatically and making the city
picker UI understand the extra level of hierarchy.
A little bit of fallout: lakeslice gridlocks again; this regression is
actually from the recent traffic signal changes, but I'm just now
regenerating everything. Will fix soon.
use the site name as the city, instead of picking the "closest" major
city. This is introducing too much friction in automation.
cyipt/actdev#65
There will be a few awkward results -- cambridge gets renamed, and lcid
gets disassociated from leeds. Worth it for now.
