Introduction

This repository contains an experimental framework with a set of baselines for the Taxi Fair Cruising problem.

Installation

Install Miniconda

Install Miniconda from https://docs.conda.io/en/latest/miniconda.html

Create a new environment by

conda create --name fair_taxi --file spec-file.txt
conda activate fair_taxi

Install Python packages

pip install -r requirements.txt

Set Up Database

Experimental framework uses MongoDB for storage of the results.

Install and set up MongoDB: https://docs.mongodb.com/manual/installation/

Install Gym Environement

Environment is distributed as a separate package.

If you want to use it without modification, install from Github by

pip install git+https://github.com/allogn/gym-taxi#egg=gym-taxi

If you want to use the development mode, use

pip install -e git+https://github.com/allogn/gym-taxi#egg=gym-taxi

Finally, if you want to contribute, fork the simulator repository, download it to some folder, and run in that folder:

python setup.py develop

This will install a package in a development mode.

Create a DAG and data directories

Create directories for data storage, and for storage of an experimental configuration files (DAGs, see #experimental-framework below).

Environment Variables

Set ALLDATA_PATH environmental variable to a directory with datasets.

export ALLDATA_PATH=<path to data>
export DAGS_PATH=<path to dags>

For automatic setting in Conda, follow this.

Experimental Framework

Experiments are defined as JSON files in DAGS_PATH directory. An example file dag_defaults.json is located in the root of the repository. Ommiting fields results with a substitution with a default value from this file.

Experiments are named by the JSON files. For example, dummy.json will produce results tagged as dummy.

Run experiments by:

python scripts/load_and_rum_experiments.py <name of the experiment>

Monitor experiments by logs in strout and by logs in tensorflow:

$ tensorflow --logdir $ALLDATA_PATH/generated/<name of experiment>

An experiment can be stopped at any moment (for example, satisfactory performance). Next run of the same experiment will continue running solvers that have never been initiated (if full_rerun parameter of an experiment is set to 0), or will clean all existing results. Rerun of a particular solver can be achieved by removing entries from the database with {"solver": "name-of-solver"}.