The meta-heuristic algorithms have aroused great attention for controller optimization. However, most of them are inseparable from the explicit system models when addressing a constrained optimization problem (COP). In this paper, we propose a data-driven constrained bat algorithm via a gradient-based depth-first search (GDFS) strategy. In the proposed scheme, the GDFS strategy can predetermine a search space that satisfies some strict constraints (e.g., stability requirements) of the optimized system. Meanwhile, an improved boundary constraint handling method is proposed to limit the exploration process to the predetermined space. In this way, the proposed algorithm can solve the COP by utilizing experimental data from real scenes, thereby relieving the dependence on precisely modeling the complex system. Together with an $\epsilon$-constraint-handling method, the bat algorithm is employed to seek the global optimum of the COP. The search performance is enhanced by the designed linear-varying elite layer-based local search and a social learning-based walk mechanism to dynamically balance exploration and exploitation. The convergence is ensured based on the criteria of the stochastic optimization algorithm. Experimental results on a servo drive system and benchmark test functions verify the effectiveness of the proposed algorithm.

元启发式算法引起了控制器优化的极大关注。然而，在解决约束优化问题 (COP) 时，它们中的大多数都与显式系统模型密不可分。在本文中，我们通过基于梯度的深度优先搜索（GDFS）策略提出了一种数据驱动的约束蝙蝠算法。在所提出的方案中，GDFS 策略可以预先确定一个满足优化系统的一些严格约束（例如，稳定性要求）的搜索空间。同时，提出了一种改进的边界约束处理方法，将探索过程限制在预定空间内。这样，所提出的算法可以利用真实场景的实验数据来解决COP，从而减轻对复杂系统精确建模的依赖。连同一个$\epsilon$-约束处理方法，采用蝙蝠算法寻求COP的全局最优值。通过设计的基于线性变化精英层的局部搜索和基于社会学习的步行机制来动态平衡探索和利用，增强了搜索性能。基于随机优化算法的标准确保收敛。伺服驱动系统的实验结果和基准测试功能验证了所提出算法的有效性。