MoCoDA: Model-based Counterfactual Data Augmentation

Authors: Silviu Pitis, Elliot Creager, Ajay Mandlekar, Animesh Garg

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that MOCODA enables RL agents to learn policies that generalize to unseen states and actions. We use MOCODA to train an offline RL agent to solve an out-of-distribution robotics manipulation task on which standard offline RL algorithms fail.
Researcher Affiliation Collaboration Silviu Pitis 1 Elliot Creager1 Ajay Mandlekar2 Animesh Garg1,2 1University of Toronto and Vector Institute, 2NVIDIA
Pseudocode No The paper describes the MOCODA framework and its steps (Figure 3), but it does not include a formally labeled "Algorithm" or "Pseudocode" block with structured steps.
Open Source Code Yes 1Visualizations & code available at https://sites.google.com/view/mocoda-neurips-22/
Open Datasets No The paper describes the empirical data used (e.g., 'empirical training data consisting of left-to-right and bottom-to-top trajectories' for 2D Navigation, and 'dataset of 50000 transitions' for Hook Sweep2), but does not provide a specific link, DOI, or formal citation for public access to these datasets themselves.
Dataset Splits Yes The models are each trained on a empirical dataset of 35000 transitions for up to 600 epochs, which is early stopped using a validation set of 5000 transitions.
Hardware Specification Yes All our experiments were run on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions using the 'Adam optimizer [26]' but does not provide specific version numbers for any programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, scikit-learn) used in the implementation.
Experiment Setup Yes We train all models for 600 epochs using the Adam optimizer [26] with a learning rate of 1e-4 and a batch size of 256.