Pith. sign in

REVIEW 2 cited by

Your Policy Regularizer is Secretly an Adversary

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2203.12592 v4 pith:XDJHIN3B submitted 2022-03-23 cs.LG stat.ML

Your Policy Regularizer is Secretly an Adversary

classification cs.LG stat.ML
keywords regularizationrewardrobustnessperturbationspolicyadversarialadversaryentropy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Policy regularization methods such as maximum entropy regularization are widely used in reinforcement learning to improve the robustness of a learned policy. In this paper, we show how this robustness arises from hedging against worst-case perturbations of the reward function, which are chosen from a limited set by an imagined adversary. Using convex duality, we characterize this robust set of adversarial reward perturbations under KL and alpha-divergence regularization, which includes Shannon and Tsallis entropy regularization as special cases. Importantly, generalization guarantees can be given within this robust set. We provide detailed discussion of the worst-case reward perturbations, and present intuitive empirical examples to illustrate this robustness and its relationship with generalization. Finally, we discuss how our analysis complements and extends previous results on adversarial reward robustness and path consistency optimality conditions.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Generalised Bellman recurrence and three dualities in sequential decision-making

    cs.LG 2026-07 conditional novelty 4.0

    Bellman's equation is derived from three compatibility conditions (D, R, A), and three known RL dualities are presented as faces of one unifying triangle.

  2. HAEPO: History-Aggregated Exploratory Policy Optimization

    cs.LG 2025-08 conditional novelty 4.0

    HAEPO weights each trajectory by its softmax-normalized cumulative log-likelihood, adds entropy and KL penalties, and matches or slightly surpasses PPO, GRPO, and DPO on small RL and summarization tasks.