Ignorance is Bliss: Robust Control via Information Gating

Authors: Manan Tomar, Riashat Islam, Matthew Taylor , Sergey Levine, Philip Bachman

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

Reproducibility Variable Result LLM Response
Research Type Experimental We apply Info Gating to various objectives such as multi-step forward and inverse dynamics models, Q-learning, and behavior cloning, highlighting how Info Gating can naturally help in discarding information not relevant for control. Results show that learning to identify and use minimal information can improve generalization in downstream tasks. Policies based on Info Gating are considerably more robust to irrelevant visual features, leading to improved pretraining and finetuning of RL models. Quantitative analyses of applying Info Gating in the context of various downstream objectives which show clear benefits in terms of improved generalization performance.
Researcher Affiliation Collaboration Manan Tomar University of Alberta Riashat Islam Mc Gill University Matthew E. Taylor University of Alberta Sergey Levine University of California, Berkeley Philip Bachman Microsoft Research Montreal
Pseudocode Yes Algorithm 1 Info Gating Pseudocode
Open Source Code No No explicit statement about open-source code availability or repository link found.
Open Datasets Yes We test 1) and 2) on the offline visual D4RL domain [25] and 3) on the Kitchen [12] manipulation domain. ... We test this version of Info Gating on CIFAR-10, while evaluating performance on Corrupted CIFAR10 [16].
Dataset Splits No The paper does not explicitly provide specific training/validation/test dataset splits (percentages, counts, or explicit standard split citations beyond dataset name).
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts) are provided for running experiments. The acknowledgment mentions 'Compute Canada' but no specific specifications.
Software Dependencies No No specific software dependencies with version numbers are provided.
Experiment Setup Yes Hyper-parameters are listed in Appendix E. ... Table 11: Visual D4RL Locomotion Training Details. Table 12: Kitchen Training Details. Table 13: CIFAR/STL-10 Training Details.