This repository contains the ongoing research on RL-powered UAV data collection from the GrADyS research group.
The repository is organized in scenarios. Each scenario presents experiments evaluating some hypothesis, each is generally an evolution of the previous. Not all scenarios generated results of value, only the ones that did will be listed in this documentation. All experiments use tensorboard for logging. The tensorboard log files contain training curves and hyperparameters used for each experiment.
Statistics that span multiple variations of the algorithm.
- The training_time and can be generated by running
statistics/training_time/compute.py. - Graphs of completion time, success rate and training time can be generated
by running
statistics/plots/plot.py.
Source and statistics about the Non-Collaborative Algorithm experiments.
Explored the applicability of MADDPG to the UAV data collection problem with scale comparable to real world scenarios. Documentation for this scenario can be found here.
Source and statistics about the Collaborative Algorithm experiments.
Explored the ability of MADDPG to learn a collaborative strategy for data collection. Documentation for this scenario can be found here.
A continuation of the previous scenario, this one explores the effects of varying the number of agents in the environment. Documentation for this scenario can be found here.
Explored the performance of an asynchronous training framework (Ape-X) for the UAV data collection problem. Documentation for this scenario can be found here.
Source and statistics about the Flexible Agent Count Algorithm and Flexible Sensor Count Algorithm experiments.
Optimization and refactoring of 003-collaboration, also including experiments with varying number of sensors and local communication. Documentation for this scenario can be found here.