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].

Your Policy Regularizer is Secretly an Adversary

Authors: Rob Brekelmans, Tim Genewein, Jordi Grau-Moya, Gregoire Detetang, Markus Kunesch, Shane Legg, Pedro A Ortega

TMLR 2022 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform experiments for a sequential grid-world task in Sec. 4 where, in contrast to previous work, we explicitly visualize the reward robustness and adversarial strategies resulting from our theory.
Researcher Affiliation Collaboration Rob Brekelmans EMAIL University of Southern California Information Sciences Institute Tim Genewein EMAIL Jordi Grau-Moya Grégoire Delétang Markus Kunesch Shane Legg Pedro Ortega EMAIL Deep Mind
Pseudocode No The paper provides mathematical formulations and derivations, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps.
Open Source Code No The paper does not contain any explicit statements about releasing source code for the described methodology, nor does it provide a link to a code repository.
Open Datasets No In Fig. 4(a), we consider a grid world where the agent receives +5 for picking up the reward pill, -1 for stepping in water, and zero reward otherwise. (This describes a custom-defined environment, not a publicly available dataset with access information.)
Dataset Splits No The paper describes experiments in a 'grid world' environment, which is a simulated task. It does not mention any training, validation, or test dataset splits, as it's not applicable in the traditional sense for this type of simulated environment.
Hardware Specification No The paper does not provide any specific details about the hardware used to run the experiments, such as GPU/CPU models, memory, or processor types.
Software Dependencies No The paper mentions 'cvx-py (Diamond & Boyd, 2016)' as a tool used, but does not provide a specific version number for cvx-py or any other key software dependencies.
Experiment Setup Yes We train an agent using tabular Q-learning and a discount factor γ = 0.99. We consider the single-step example in Sec. 4.1 Fig. 2 or App. H Fig. 10-11, with a two-dimensional action space, optimal state-action value estimates, Q (a, s) = r(a, s) = {1.1, 0.8}, and uniform prior π0(a|s). The case of policy regularization with α = 2 and β = 10 is particularly interesting.