Notes for 2016-09-15 morning

[cathywu]

Todo

• Review the PR (new sweeps and new controllers!).
• Take a look at the TrafficJam plots from vtype_traffic_jam.py: https://www.dropbox.com/sh/0skbb3ij4ee30j1/AACQPqWRNXBdGHHM3jafsdXqa?dl=0. Seems promising in terms of energy. Our controllers are in general more efficient than IDM. The human behavior is very efficient but also very slow.
• Create a more massive human jam.
  • We should be able to recreate the Japanese (Sugiyama) loop road backwards propagating waves. This will serve as the baseline for the experiments. Maybe all of our experiments can be based on these parameters (230 meter loop, 22 vehicles).
  • I've set up the single-lane version (see sugiyama_jam.py); can you confirm that there are backwards propagating waves?
  • Here are plots for various controllers: https://www.dropbox.com/sh/gtyd3wsul26e75v/AABl6rSOMQPpWU32s0yTL4xAa?dl=0. Summary: IDM and ACC do worse than humans; Midpoint does slightly better. This is mostly good news!
• Relatedly, in the speed profile, can we also plot the 25th and 75th percentile speeds? This may help us determine if the following figure, for example, has backwards propagating waves:
  
• Let's set up the 2-lane version of the Sugiyama Jam. The experiment is in sugiyama_jam.py. When I set numLanes=2, I get collisions errors in initializing the cars. Are they all being initialized on the same lane or something? Sample errors:

Warning: Vehicle 'human-001' performs emergency stop at the end of lane 'top_0' for unknown reasons (decel=0.00), time=0.00.
Warning: Teleporting vehicle 'human-005'; collision with vehicle 'human-000', lane='right_0', gap=-0.10, time=0.00 stage=move.
Warning: Vehicle 'human-005' ends teleporting on edge 'top', time 0.00.

• Plot the 2-lane behavior of the Sugiyama setting for each of our controllers, similarly to what I did above with the 1-lane version.
• Tweak the lane changing behavior for the human agents, which results in more instability. Based on what I've seen so far, I expect/hope that the controllers (fillgapmidpoint, in particular, and perhaps also midpoint) will be able to smooth out some of this instability. Fingers crossed.
• Then experiment with a small number of non-human controllers (similar to few_robots_sweep.py), with the total still adding up to 22*numLanes. Unfortunately, based on what I've seen so far, I don't expect this "few robots" experiment to yield good results. I think we need a good fraction of the vehicles to be controlled in order to achieve a speed increase, for instance. I have not examined the effect on energy though, so that could still be a positive result.

Questions

• What is the purpose of the color scale on plots in the lower right-hand corner? See example below.
• Do we have different colors for different vehicle types? Also see example below (all points appear red).
• The energy plots are for all vehicles, not just robot ones, right?