In dynamic industrial environments, agents are required to continuously adapt to evolving data streams and distribution shifts. Existing online optimization methods for sequential decision-making typically focus on maximizing immediate rewards, while lacking explicit modeling of temporal dependencies between policy outputs at adjacent time steps. This may lead to high-frequency fluctuations in consecutive actions and consequently unstable performance. To address this issue, this paper proposes a real-time decision optimization algorithm termed the Temporal Consistency Constrained Agent (TCC-Agent). Within a policy gradient framework, a temporal consistency constraint is incorporated into the objective function by penalizing discrepancies between consecutive actions as well as differences in policy distributions across adjacent time steps. In this way, temporal smoothness is integrated into the joint loss, enabling more stable regulation of the policy update process. The proposed framework consists of a policy network and a consistency constraint module. The constraint term directly participates in loss computation without introducing additional learnable parameters. Experimental results on continuous control benchmarks and an industrial load simulation environment demonstrate that the proposed method achieves improvements in policy stability and loss fluctuation control compared with baseline algorithms. Specifically, it reduces loss variance and action variation rates, while maintaining more stable convergence under non-stationary conditions. These results indicate that incorporating temporal consistency constraints can effectively mitigate high-frequency policy variations and enhance the robustness and stability of sequential decision-making processes.
首先,在机制融合层面,多数方法将时序一致性作为一种辅助性的正则化手段。例如,Mysore等[2]提出的CAPS(Conditioning for Action Policy Smoothness)方法通过在动作输出层施加平滑性正则抑制振荡。而Kim[3]的工作则侧重于约束Q函数在相邻时间步的变化。这类方法将一致性作为辅助损失项,并未将其深度耦合至策略优化的核心梯度更新逻辑中。这种“外围式”的约束在高频动态环境下往往显得响应迟缓,难以在实时响应性与策略平滑性之间实现动态、自适应的平衡,从而导致收敛过程不稳定或决策延迟。
ZHAOY, ZHAOW S, BONEYR, et al. Simplified temporal consistency reinforcement learning[C]//Proceedings of the 40th International Conference on Machine Learning. Cambridge :ML Research Press,2023: 42227-42246. DOI:10.5555/3618408.3620184 .
[2]
MYSORES, MABSOUTB, MANCUSOR, et al. Regularizing action policies for smooth control with reinforcement learning[C]//2021 IEEE International Conference on Robotics and Automation (ICRA). New York: IEEE Press, 2021: 1810-1816. DOI:10.1109/ICRA48506.2021.9561138 .
[3]
KIMC. Temporal consistency-based loss function for both deep Q-networks and deep deterministic policy gradients for continuous actions[J]. Symmetry, 2021, 13(12): 2411. DOI:10.3390/sym13122411 .
[4]
KOBAYASHI T. L2C2: Locally Lipschitz continuous constraint towards stable and smooth reinforcement learning[C]//2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). New York: IEEE Press, 2022: 4032-4039. DOI:10.1109/IROS47612.2022.9981812 .
TODOROVE, EREZT, TASSAY. MuJoCo: A physics engine for model-based control[C]//2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. New York: IEEE Press, 2012: 5026-5033. DOI:10.1109/IROS.2012.6386109 .
[7]
MCINROET, SCHÄFERL, ALBRECHTS V. Learning temporally-consistent representations for data-efficient reinforcement learning[DB/OL]. [2025-04-02]. DOI: 10.48550/arXiv.2110.04935 .
[8]
ZHANGD, CAIM, LIJ, et al. TDRM: Smooth reward models with temporal difference for LLM RL and inference[DB/OL]. [2025-04-02].
[9]
ZHANGJ, FANG LINJ J, SHIK X, et al. Q-Distribution guided Q-learning for offline reinforcement learning: Uncertainty penalized Q-value via consistency model[DB/OL]. [2025-04-02]. DOI: 10.52202/079017-1724 .
[10]
GAOC Q, XUK, ZHOUK Q, et al. Value penalized Q-learning for recommender systems[DB/OL]. [2025-04-02]. DOI: 10.1145/3477495.3531796 .
[11]
THODOROFFP, DURANDA, PINEAUJ, et al. Temporal regularization for Markov decision process[DB/OL]. [2025-04-02]. DOI: 10.48550/arXiv.1811.00429 .
[12]
SHENQ L, LIY, JIANGH M, et al. Deep reinforcement learning with robust and smooth policy[DB/OL]. [2025-04-02].
[13]
SONGX, DUANJ, WANGW, et al. LipsNet: A smooth and robust neural network with adaptive Lipschitz constant for high accuracy optimal control[DB/OL]. [2025-04-02].
[14]
PARANJAPEA, QUADERN, UHLMANNL, et al. Reinforcement learning agent for multi-objective online process parameter optimization of manufacturing processes[J]. Applied Sciences, 2025, 15(13): 7279. DOI:10.3390/app15137279 .
[15]
LID Z, GUW T, SONGT H. Multi-objective reinforcement learning in process control: A goal-oriented approach with adaptive thresholds[J]. Journal of Process Control, 2023, 129: 103063. DOI:10.1016/j.jprocont.2023.103063 .
[16]
SCHÄFERG, SELIGERR, REHRLJ, et al. Multi-objective reinforcement learning for energy-efficient industrial control[C]//Database and Expert Systems Applications - DEXA 2025 Workshops. Cham: Springer, 2026: 67-72. DOI:10.1007/978-3-032-02003-1_6 .
[17]
KLARM, SCHWORMP, WUX Q, et al. Transferable multi-objective factory layout planning using simulation-based deep reinforcement learning[J]. Journal of Manufacturing Systems, 2024, 74: 487-511. DOI:10.1016/j.jmsy.2024.04.007 .
KUZNETSOVA, SHVECHIKOVP, GRISHINA, et al. Controlling overestimation bias with truncated mixture of continuous distributional quantile critics[DB/OL]. [2025-04-02].
[24]
JANUSZEWSKIP, OLKOM, KRÓLIKOWSKIM, et al. Continuous control with ensemble deep deterministic policy gradients[DB/OL]. [2025-04-02]. DOI: 10.48550/arXiv.2111.15382 .
[25]
HONARIH, SOUFI ENAYATIA M, GHAFARIAN TAMIZIM, et al. Meta SAC-lag: Towards deployable safe reinforcement learning via MetaGradient-based hyperparameter tuning[C]//2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). New York: IEEE Press, 2024: 619-626. DOI:10.1109/IROS58592.2024.10802547 .