Three New Papers Published This Week

JAIR’s October 22 issue contains three new studies. The first introduces Prob‑IRM, a method for learning reward machines (RMs) when sensor data is noisy. It uses inductive logic programming to infer RM rules from noisy proposition histories and interleaves RM learning with policy optimisation. Experiments show it performs comparably to baselines using handcrafted RMs in both discrete and continuous environments.¹

The second paper presents an empirical Bayesian framework for contextual bandits. Jiang and Zhu estimate the covariance matrix of the prior distribution and embed it into two algorithms, ebmTS and ebmUCB. They prove frequentist regret bounds and demonstrate up to 15 % higher click‑through rates in simulations compared to standard methods. Further testing will be needed before deployment.¹

The third study surveys explainable reinforcement learning (XRL). Saulières reviews over 250 papers, proposes a taxonomy based on “What” (the target explained) and “How” (the explanation mechanism), and highlights emerging needs such as counterfactual explanations. EU legislation that requires explainability for high‑risk AI applications takes effect January 1, 2027, making this survey timely.¹

Expert context: Prob‑IRM addresses the limitation that traditional RM learning assumes perfect state mapping; noise can degrade RL performance by even small amounts. The empirical Bayesian approach tackles task dependencies often ignored by other multi‑bandit algorithms, potentially improving decision making across related tasks. The XRL taxonomy aligns with regulatory pressure for transparency and may guide future research toward more comprehensive coverage.