SOCS-Project

Run

bash run_simulation.sh ${paramset}

to get a test flight running. (Try for example paramset=11)

This will show show both the graphs for visualization and a 3D model of a drone (Shoutout to Jan Schiffeler, @Platygator, for the visualization).

Alternatively, to have a slightly quicker simulation but without the 3D model, run

bash run_simulation_without_viz.sh ${paramset}

This will start a simulation where a drone collects coins successively in predefined positions. Each coin will earn the drone a reward, the reward score will decrease with each passed time step.

As of now, there is no real use of the rewards system. Next steps would be to implement a training routine to make use of the reward.

Neuroevolution

Training

Executing

python evolve/evolution.py

starts a simple neuroevolution algorithm to train and evolve the controller unit within the drones of a population.

Current objective is to collect a coin and/or stay in the air for as long as possible.

Status Quo: Training is enabled, but the training is not very successful as of yet. Some things can be improved, such as:

• Parallelization of fly() function
• Training strategy:
  • Tune hyperparameters such as mutation rate, crossover rate, ...
  • Tune rewards for time in air, distance to coin, coin collected, crashed, and so on

Result

To see the best drone of a certain training run in action, execute

bash show_best_drone.sh ${execution_time}

Test Framework

You can write unit tests for your code by placing files of test classes in the ./test directory.

You can take test/test_sensor.py as an example for the correct syntax.

Then simply execute the script run_tests.sh (or manually execute python -m pytest . --cov ./py --cov-report html)