Machine learning-guided conflict selections for Conflict-Based Search.

Follow instructions in CBSH2 to compile the c++ code.

A set of instances for the lak503d map are in the instance folder.

Below is an example to collect data for the lak503d map for the instance instances/lak503dmap-100agents-0.agents with 2 hours runtime. The option -u 1 specifies data collection mode.

./CBSH2 -m instances/lak503d.map -o collect_log.csv -t 7200 -s 1 -h WDG -a instances/lak503dmap-100agents-1.agents -u 1 --feature feature/feature_Lak503d_0.txt

After data collection, process the data with the following python script for it to be compatible with the SVM rank software.

python dataProcessor.py

After that, one can use the SVM rank software to train the model using the produced data files.

Once finished training, there will be a trained SVM rank model the main folder. Open the file and extract the weights into a new weight file weight_file with the content being rows of pairs featureID weights.

To run test on the lak503d map with 1 hour runtime on instance instances/lak503dmap-100agents-2.agents, one can use the command below. The option -u 4 specifies test mode.

./CBSH2 -m city/Paris/lak503d.map -o run_test_ML.csv -t 3600 -s 1 -h WDG -a instances/lak503dmap-100agents-2.agents -k 100 -u 4  --model weight_file

Please cite our paper if you use this code in your work.

@inproceedings{huang2021conflict,
  title={Learning to resolve conflicts for multi-agent path finding with conflict-based search},
  author={Huang, Taoan and Dilkina, Bistra and Koenig, Sven},
  booktitle={AAAI Conference on Artificial Intelligence},
  year={2021},
  pages={11246--11253}
}