Han Zhang, Zihan Gu, Zhiyuan Wang, Tianyi Ma, Jiacheng Lu, Xinyan Zhang, Yuhao Wei, Cheng Hua. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Interpretable reinforcement learning policies are essential for high-stakes decision-making, yet optimizing decision tree policies in Markov Decision Processes (MDPs) remains challenging. We propose SPOT, a novel method for computing decision tree policies, which formulates the optimization problem as a mixed-integer linear program (MILP). To enhance efficiency, we employ a reduced-space branch-and-bound approach that decouples the MDP dynamics from tree-structure constraints, enabling efficient parallel search. This significantly improves runtime and scalability compared to previous methods. Our approach ensures that each iteration yields the optimal decision tree. Experimental results on standard benchmarks demonstrate that SPOT achieves substantial speedup and scales to larger MDPs with a significantly higher number of states. The resulting decision tree policies are interpretable and compact, maintaining transparency without compromising performance. These results demonstrate that our approach simultaneously achieves interpretability and scalability, delivering high-quality policies an order of magnitude faster than existing approaches.

ABSTRACT The zeroing neural network (ZNN), a canonical recurrent neural network, was developed in previous studies to address time‐varying problem‐solving scenarios. Numerous practical applications involve time‐varying linear equations and inequality systems that demand real‐time solutions. This article proposes a ZNN model specifically designed to solve such time‐varying linear systems. Innovatively, it incorporates a new non‐negative slack variable that transforms complex time‐varying inequality systems into more easily solvable time‐varying equation systems. By using an exponential decay formula and establishing an indefinite error function, the ZNN model is built. The suggested ZNN model's convergence properties are validated by theoretical research. Results from comparative simulations further support the superiority and effectiveness of the ZNN model in resolving inequality systems and time‐varying linear equations.

ABSTRACT This paper proposes a robust dual‐integral structure zeroing neural network (ZNN) design framework, effectively overcoming the limitations of existing single‐integral enhanced ZNN models in completely suppressing linear noise. Based on this design framework, a complex‐type dual‐integral structure ZNN (DISZNN) model with inherent linear noise suppression capability is constructed for computing dynamic complex matrix inversion (DCMI) online. The stability, convergence, and robustness of the proposed DISZNN model are ensured via rigorous theoretical analyses. In three distinct experiments involving DCMI (including cases with only imaginary parts, both real and imaginary parts, and high‐dimensional scenarios), the state trajectories of the DISZNN model are well and quickly fitted to the dynamic trajectories of the theoretical solutions with very low residual errors in various linear noise environments. More specifically, the residual errors of the DISZNN model for online computation of DCMI under linear noise environments are consistently below the order of , representing one‐thousandth of the residual errors in existing noise‐tolerant ZNN models. Finally, the DISZNN design framework is applied to construct a controlled chaotic system of a permanent magnet synchronous motor (PMSM) with uncertainties and external disturbances based on real‐world modeling. Experimental results demonstrate that the three state errors of the controlled PMSM chaotic system converge to zero quickly and stably under various conditions (system parameters, external disturbances, and uncertainties), further highlighting the superiority and generalizability of the DISZNN design framework.

With the rapid development of the multi-robot systems, formation control has become a fundamental challenge. Traditional approaches focus mainly on the design of control algorithms to realize specific formation patterns, while neglecting how to determine the desired formation. In this paper, the optimal formation problem based on shape theory is reformulated as a convex optimization problem. A predetermined time convergent zeroing neural dynamics (PDTZND) approach, derived from zeroing neural networks (ZNN), is proposed to efficiently solve this problem. The PDTZND approach ensures that the system error converges in a strict and predetermined time, which provides an efficient, accurate solution for optimal formation. In addition, the convergence of the proposed approach is rigorously analyzed by means of Lyapunov theory, and its validity and superiority are verified by numerical simulations and physical experiments.

Abstract Zeroing neural networks (ZNNs), a specialized class of recurrent neural networks, have demonstrated remarkable effectiveness in matrix computation and dynamic optimization problems due to their inherent parallel computing capabilities. In this paper, a predefined-time and noise reduction ZNN (PTNRZNN) model is proposed for solving convex quadratic programming problems with equality and inequality constraints. Additionally, a new activation function is proposed, demonstrating enhanced accelerated convergence and noise reduction performance compared to previous models. The convergence and robustness of the PTNRZNN model are effectively proven through theoretical assessment. Furthermore, the performance of the PTNRZNN model is further validated through simulation experiments. Finally, the PTNRZNN model is applied to the binary assignment problem in logistics (BAPL), yielding optimized results with an error margin as low as $$10^{-2}$$ 10 - 2 compared to theoretical values. The strong robustness of the method makes it an excellent performer in solving BAPL under noise interference.

In recent years, some new generators of binary sequences, such as the clock-controlled shift register and the cascade-connected clock-controlled shift register, have been suggested. Most sequences generated by these models have period of the form (2/sup n/-1)/sup k/. Further, many other kinds of binary sequences have this kind of period. Here, the authors give the lower bound of linear complexity of all these kinds of sequences that have period of the form (2/sup n/-1)/sup k/ with n being a prime.>