Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

On Minimizing Adversarial Counterfactual Error in Adversarial Reinforcement Learning

Authors: Roman Belaire, Arunesh Sinha, Pradeep Varakantham

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical evaluations on standard benchmarks (Mu Jo Co, Atari, and Highway) demonstrate that our method significantly outperforms current state-of-the-art approaches for addressing adversarial RL challenges, offering a promising direction for improving robustness in DRL under adversarial conditions.
Researcher Affiliation Academia Roman Belaire Singapore Management University Singapore EMAIL Arunesh Sinha Rutgers University New Brunswick, NJ EMAIL Pradeep Varakantham Singapore Management University Singapore EMAIL
Pseudocode Yes Algorithm 1: δ-PPO
Open Source Code Yes Our code is available at https://github.com/romanbelaire/acoe-robust-rl.
Open Datasets Yes Our empirical evaluations on standard benchmarks (Mu Jo Co, Atari, and Highway) demonstrate that our method significantly outperforms current state-of-the-art approaches
Dataset Splits No We report the mean result over 5 policies initialized with random seeds, with 50 test episodes each.
Hardware Specification Yes We train our linear models on an NVIDIA Tesla V100 with 16gb of memory, and LSTM models on an NVIDIA L40 32gb GPU.
Software Dependencies No The paper does not provide specific software dependencies with version numbers. It mentions algorithms and frameworks like PPO, DQN, Adam, and LSTM, but not their specific software implementations or versions.
Experiment Setup Yes We train our methods for 900 episodes for all Mu Jo Co environments, using an annealed (Adam) learning rate of 0.005. The robustness hyperparameter λ is set to 0.2 for all of our models, which is the same as the robustness hyperparameters found in prior works Oikarinen et al. (2021); Liang et al. (2022); Belaire et al. (2024); Zhang et al. (2020). The attack neighborhood sample size is set to 10, and the training attack neighborhood radius is set to ϵ = 0.1, both tuned from sets in the range 100%.