English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Reinforcement Learning (RL) algorithms have led to recent successes insolving complex games, such as Atari or Starcraft, and to a huge impact inreal-world applications, such as cybersecurity or autonomous driving. In theside of the drawbacks, recent works have shown how the performance of RLalgorithms decreases under the influence of soft changes in the rewardfunction. However, little work has been done about how sensitive thesedisturbances are depending on the aggressiveness of the attack and the learningexploration strategy. In this paper, we propose to fill this gap in theliterature analyzing the effects of different attack strategies based on rewardperturbations, and studying the effect in the learner depending on itsexploration strategy. In order to explain all the behaviors, we choose asub-class of MDPs: episodic, stochastic goal-only-rewards MDPs, and inparticular, an intelligible grid domain as a benchmark. In this domain, wedemonstrate that smoothly crafting adversarial rewards are able to mislead thelearner, and that using low exploration probability values, the policy learnedis more robust to corrupt rewards. Finally, in the proposed learning scenario,a counterintuitive result arises: attacking at each learning episode is thelowest cost attack strategy.

Disturbing Reinforcement Learning Agents with Corrupted Rewards