Goal-Conditioned Predictive Coding for Offline Reinforcement Learning

Authors: Zilai Zeng, Ce Zhang, Shijie Wang, Chen Sun

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive empirical evaluations on Ant Maze, Franka Kitchen and Locomotion environments, we observe that sequence modeling can have a significant impact on challenging decision making tasks.
Researcher Affiliation Academia Zilai Zeng Brown University Ce Zhang Brown University Shijie Wang Brown University Chen Sun Brown University
Pseudocode Yes Appendix B Pseudocode of GCPC. Algorithm 1 Goal-conditioned Predictive Coding (GCPC) for Rv S
Open Source Code Yes Our code is available at https://brown-palm.github.io/GCPC/.
Open Datasets Yes To answer the questions above, we conduct extensive experiments on three domains from D4RL offline benchmark suite [16]: Ant Maze, Franka Kitchen and Gym Locomotion.
Dataset Splits Yes During policy learning, we use the pre-trained Traj Net that achieves the lowest validation reconstruction loss to generate the bottleneck.
Hardware Specification Yes All experiments are performed on a single Nvidia RTX A5000.
Software Dependencies No The paper mentions using Adam optimizer but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes Appendix A.3 Hyperparameter. We list hyperparameters for BC and Rv S-G/R replication in Table A1 and GCPC implementation in Table A2. For Rv S-G/R and Policy Net in GCPC, we use a two-layer feedforward MLP as the policy network, taking the current state and goal (state or return-to-go) as input, the only difference is that Policy Net takes the bottleneck as an additional input. All experiments are performed on a single Nvidia RTX A5000.