Latent exploration for Reinforcement Learning

Authors: Alberto Silvio Chiappa, Alessandro Marin Vargas, Ann Huang, Alexander Mathis

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

Reproducibility Variable Result LLM Response
Research Type Experimental With extensive experiments, we show that Lattice can replace standard unstructured exploration [2, 5] and time-only-correlated exploration (g SDE) [8] in off-policy (SAC) and on-policy (PPO) RL algorithms, and improve performance in complex motor control tasks. Importantly, we demonstrate that Lattice-SAC is competitive in standard benchmarks for continuous control, such as the locomotion environments of Py Bullet [20]. We benchmarked Lattice on standard locomotion tasks [47, 6, 16, 48 50] in Py Bullet [20], as well as musculoskeletal control tasks of Myo Suite [18] built in Mu Jo Co [31]. All the results are averaged across 5 random seeds.
Researcher Affiliation Academia Alberto Silvio Chiappa École Polytechnique Fédérale de Lausanne (EPFL) alberto.chiappa@epfl.ch Alessandro Marin Vargas EPFL alessandro.marinvargas@epfl.ch Ann Zixiang Huang Mila, EPFL zixiang.huang@mail.mcgill.ca Alexander Mathis EPFL alexander.mathis@epfl.ch
Pseudocode Yes Algorithm 1 Standard (e.g., PPO, SAC) and Algorithm 2 Lattice are presented on page 4.
Open Source Code Yes The code is available at: https://github.com/amathislab/lattice.
Open Datasets Yes We benchmarked Lattice on standard locomotion tasks [47, 6, 16, 48 50] in Py Bullet [20], as well as musculoskeletal control tasks of Myo Suite [18] built in Mu Jo Co [31].
Dataset Splits No The paper does not explicitly provide training/validation/test dataset splits with percentages or sample counts. It refers to training in environments and averaging results over random seeds, which is common in RL, but not explicit dataset splits.
Hardware Specification No The training was run on a GPU cluster, for a total of approximately 10,000 GPU-hours. No specific GPU models, CPU types, or detailed cluster specifications are provided beyond 'GPU cluster'.
Software Dependencies No We implemented Lattice as an extension of g SDE in the RL library Stable Baselines 3 [45]. While Stable Baselines 3 is mentioned, no specific version number is provided for it or any other software dependency.
Experiment Setup Yes We used the same network architecture and hyperparameters for SAC specified in [46] for all the environments (see Appendix A.7). Tables T3, T4, and T5 in Appendix A.7 provide detailed hyperparameters for SAC, PPO, g SDE, and Lattice for various tasks.