This is the code base for our paper on [Revisiting Bisimulation Metric for Robust Representation in Reinforcement Learning]. Our method introduces a new metric for measuring state-action pair similarity with adaptive coefficients, overcoming the limitations of traditional bisimulation metrics.
Our experimental setup is based on the DrQ-v2 framework. For details on dependencies and environment configuration, please refer to the environment.yml file.
To train the model, use the following scripts:
# Run on a cluster
$ bash cluster_run.sh
# Run locally
$ bash run.sh
Ensure to update the scripts if needed to match your specific setup.
For empirical results and further details, refer to the paper. Our experiments demonstrate significant improvements in sampling efficiency and task performance.