flow-project / flow / 3884

89%

push

travis-ci-com

Multi-agent traffic light grid (#623)

New multi-agent environment.

## Description

- Introduces multi-agent version of `PO_TrafficLightsGridEnv`.
- Includes some updated, revised, and cleaned up documentation relating to multi-agent Flow.
- Includes parameterized run script, which supports different grid sizes, local vs cluster runs, different vehicle inflow rates.

* revised multiagent documentation

* PEP8 and documentation for PO_TrafficLightGridEnv

* Add multiagent version of PO_TrafficLightGridEnv, including helper function for retreiving relative nodes

* fix: observations should include num_observed vehicles from each incoming lane

* fix conversion to edge numbers

* prepare run script

* cleaning

* cleaning

* documentation

* PEP8

* add ES to multi-agent run script

* missed some multi-agent params

* probably can't use PPO policy graph for ES

* ES --> PPO (multi-agent not supported, it seems)

* bigger experiment, documentation

* clean up docs

* PEP8

* pydocstyle errors

* Update multiagent_traffic_light_grid.py

PEP8

* keep track of observed vehicles

* add unit tests for get_relative_node

* generalize multiagent grid run script

* flake8

* fix unit tests for grid

* flake8

* add inflow rate to arguments for multiagent grid script; expand parameter tuning grid search; minor tweaks

* flake8

* add local vs cluster run mode

* normalize per agent rewards

* num edges fix

* run script tweaks

* multiagent grid test

* disable autoscaling

* refactor multi-agent grid run script

* small fixes and tweaks

* flake8

* flake8

* fix unit test for multiagent traffic light grid

* docs

* docs

* remove ES

* flake8

* exception handling

* partial addressing of Eugene's comments

* addressing remainder of Eugene's comments

234 of 234 new or added lines in 6 files covered. (100.0%)

8415 of 9292 relevant lines covered (90.56%)

0.91 hits per line