This work was done in collaboration of AIRI, DeepPavlov.ai, and London Institute for Mathematical Sciences.
Create a conda environment from the export file:
conda env create -f srmt_env_export.ymlSRMT training is done with the train.py script.
To modify the default training configuration values, one can use the command line arguments corresponding to the names of variables in srmt/training_config.py.
For example, SRMT with reward functions from the paper is trained as follows.
Sparse reward function:
python train.py --experiment=<name of the folder to store checkpoints> --attn_core=true --use_rnn=false --core_memory=true --const_reward=true --intrinsic_target_reward=0 --seed=<random seed>Dense reward function:
python3 train.py --experiment=<name of the folder to store checkpoints> --attn_core=true --use_rnn=false --core_memory=true --const_reward=true --seed=<random seed>Moving Negative reward function:
python3 train.py --experiment=<name of the folder to store checkpoints> --attn_core=true --use_rnn=false --core_memory=true --any_move_reward=true --seed=<random seed>Directional reward function:
python3 train.py --experiment=<name of the folder to store checkpoints> --attn_core=true --use_rnn=false --core_memory=true --target_reward=true --positive_reward=true --intrinsic_target_reward=0.005 --seed=<random seed>Directional Negative reward function:
python3 train.py --experiment=<name of the folder to store checkpoints> --attn_core=true --use_rnn=false --core_memory=true --target_reward=true --reversed_reward=true --seed=<random seed>To evaluate the trained model on the test set of environments, use:
python eval.pyTo run a single episode with the trained SRMT agents and produce an animation:
python3 example.pyThe animation will be stored in the folder containing the experiment checnkpoints.
To avoid performance issues, it is recommended to set the following environment variables restricting Numpy CPU threads to 1:
export OMP_NUM_THREADS="1"
export MKL_NUM_THREADS="1"
export OPENBLAS_NUM_THREADS="1"@misc{sagirova2025srmtsharedmemorymultiagent,
title={SRMT: Shared Memory for Multi-agent Lifelong Pathfinding},
author={Alsu Sagirova and Yuri Kuratov and Mikhail Burtsev},
year={2025},
eprint={2501.13200},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2501.13200},
}
The repository is inspired by the Learn to Follow repository.
