Papers for
automated control system designers
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Reinforcement learning improves with smart reward label choices
Uncertainty-Guided Reward Labeling for Reinforcement Learning under Limited Feedback
Abstract: In reinforcement learning from limited feedback (RLLF), only a small fraction of an offline dataset can be labeled with rewards, and the central question is which samples should be labeled to learn a high-performing policy from the resulting partially labeled dataset. Prior work formalized this as a reward-selection problem by focusing on the selection stage while treating downstream policy learning as a black box, in a regime where queried rewards are not retained for reward-model training. We instead study the retained-label setting, where queried rewards can be stored and used to fit a reward model before policy learning. We bound the suboptimality of the learned policy in terms of two sources of error: one arising from offline RL on a fixed, offline dataset, and the other from uncertainty in the reward model. Since reward selection cannot change the offline dataset, the limited labeling budget must be used to strategically reduce reward uncertainty. Motivated by RLLF's observation that useful rewards tend to keep the agent on high-return trajectories, we introduce Successor-Guided Uncertainty REduction (SURE), a novel method that uses successor features to select rewards that are both uncertain and associated with states from which high-valued states are reachable. We formally derive SURE from a bound-induced design objective and characterize its exact one-step marginal gain. Empirically, SURE reaches near full-feedback performance with few reward labels across a variety of domains and offline-RL backbones, yielding an effective method for feedback-efficient reinforcement learning.
Finite controllers face limits in achieving exact game values under strict rules
Attaining Optimal Values under Sure Constraints
Abstract: A game may have, for each target below its value, a finite controller meeting that target, yet no finite controller attaining the value itself. We study finite stochastic games with a hard specification that must hold on every supported execution, including paths of probability zero. Max uses a pure finite controller; Min may use any pure public-history strategy. We distinguish three numerical questions: can a supplied target be met, can the unspecified optimum be attained, and what is one exact value? For explicit hard and Boolean soft Streett conditions, the strict and non-strict supplied-threshold problems are coNP-complete, attainment of the unsupplied supremum is $\mathrm P^{\mathrm{NP}}$-complete, and one exact reduced fraction is $\mathrm{FP}^{\mathrm{NP}}$-complete under polynomial metric reductions. The proofs preserve both the value and its finite attainability when fixing a single opponent, and encode adaptive queries without computing unknown subgame values. The same construction yields matching value-comparison and exact-output lower bounds for ordinary stochastic Streett games with unrestricted randomized history strategies. We also bound the memory needed for optimal finite implementations for arbitrary Muller hard conditions and rational rewards in $[0,1]$ determined by infinitely recurring states. One controller is sure from the entire original sure region and attains at every finitely attained start, with no attainment premise on the other states. For $n$ states in that region it uses $2^{O(n\log(n+1))}Q(1+\log R)$ modes, where $Q$ is the product of the least common denominators of full original chance rows sourced in that sure region and $R$ clears all reward values. Its joint dependence on $Q$ and $R$ is worst-case sharp at each fixed $n\ge6$ when $Q,R\ge2$. Optimal dependence on arena size and efficient general synthesis are not resolved here.
Discovery system improves experiments by changing theory mechanisms
Learning to Outgrow a Theory: Experimental Discovery Beyond the Initial Hypothesis Space
Abstract: Scientific discovery systems typically optimize experiments within a fixed hypothesis space. This creates a failure mode when all available candidates omit the same missing mechanism: candidate disagreement can collapse even while the model class is systematically wrong. We formulate experimental model-class revision, in which a discovery policy jointly proposes a structural edit and a diagnostic experiment that tests whether that edit is necessary. The method couples a class-level distinguishability objective, in which one shared parameterization must explain all selected experiments, with anytime-valid sequential evidence that triggers structural revision only after the current class is rejected. On 400 held-out controlled dynamical environments, the joint policy reaches 89.5% exact recovery with a budget of 32 real experiments, improving the strongest matched baseline by 10.0 percentage points while requiring fewer executed experiments and candidate fits. The learned revision-experiment pairing transfers across unseen mechanism combinations, held-out but expressible primitives, parameter extrapolation, and shifted experiment costs; when the true mechanism is outside the edit grammar, it detects library insufficiency in 88% of cases with a 5.5% false-support rate. Revision gains also transfer to ODEBench and ODEBase model-library tasks, as well as DiscoverPhysics worlds. These results support a view of scientific discovery in which deciding what mechanisms a theory should make expressible and where to collect evidence are treated as a single sequential decision problem.