ABSTRACT Currently, the cross‐chain field lacks a systematic performance evaluation framework that combines theoretical rigor with engineering scalability. Existing models struggle to accommodate multilevel serial architectures based on relay mechanisms and often overlook various engineering constraints and operational anomalies. This results in significant discrepancies between theoretical findings and actual operational characteristics. To address the aforementioned research gap, this article selects BitXHub, a representative distributed relay‐based cross‐chain framework, as the subject of study and constructs a two‐stage tandem network performance analytical model based on queuing theory. Key innovations include: (1) Incorporating real‐world cross‐chain anomalies, such as consensus failures and transaction retransmissions, to adjust service rates (modeled as exponential distributions) and mitigate errors arising from the idealized assumptions of traditional models. (2) Characterizing the number of transactions per block using a stochastic distribution rather than a fixed block size, thereby aligning the model with the system's mechanism of batch‐producing block at scheduled intervals. (3) Decomposing BitXHub's relay operations into a two‐stage subsystem and deriving key performance metrics, such as transaction latency and throughput—by combining the matrix geometric method with the Laplace–Stieltjes transform. (4) Establishing an experimental testbed based on the open source BitXHub platform and validating the model using actual operational parameters. Experimental results demonstrate that block size, transaction pool limits, and transaction load are the primary factors determining cross‐chain throughput and latency. The system exhibits robustness against service rate fluctuations under light to moderate loads, whereas performance volatility increases significantly as the load approaches the saturation threshold.
ABSTRACT Predictive autoscaling is intended to reduce the delay inherent in reactive resource control, but improvements over the Kubernetes horizontal pod autoscaler (HPA) may arise from differences in scaling signal, capacity estimation, and control policy rather than from forecasting alone. This study examines that distinction on an eight‐service polyglot Kubernetes application implemented in Go, Java/Spring Boot, and Node.js. Residual long short‐term memory (LSTM) and gated recurrent unit (GRU) forecasters are compared with policy‐matched naive persistence and linear extrapolation controllers that use the same telemetry, capacity parameters, decision interval, replica bounds, and cooldown rules. HPA, Static, and Prophet are included as additional references. The evaluation comprises 160 benchmark runs across saturation, burst, ramp, and subsaturation workloads. Under the 400‐VU subsaturation workload, GRU and LSTM reduce client p95 latency by 15.8% and 15.1% relative to HPA, respectively, without a detectable increase in replica‐hours, although both operate at higher CPU utilization. Under 1000‐VU saturation, latency comparisons are confounded by timeout truncation, while GRU and LSTM use 17.5% and 15.4% fewer replica‐hours than HPA. No significant advantage is detected for GRU over naive persistence, whereas GRU outperforms linear extrapolation on client p95 and median latency. The policy‐matched controllers also exhibit similar first‐scale timing. Overall, the results indicate that the observed advantage over HPA is associated with the complete controller design rather than learned forecasting alone, while predictor choice can still affect performance under some workload conditions.
ABSTRACT Achieving efficient path planning for unmanned underwater vehicles in underwater environments with multiple obstacles and unpredictable conditions remains a challenge. The problems of low exploration efficiency and long training time for deep reinforcement learning‐based path planning algorithms have not yet been well addressed, especially in underwater environments. Additionally, the fixed‐interval target network update mechanism is prone to causing policy oscillation during the later stages of training. To address these issues, this paper proposes a regional firefly algorithm‐enhanced Dueling Double Deep Q‐Network algorithm. This method leverages the rapid optimization capability of the firefly algorithm for auxiliary exploration, reducing the frequency of random movements by the D3QN algorithm during the early exploration phase to accelerate its convergence. Meanwhile, an adaptive target update strategy is designed to dynamically adjust the update frequency of the target network based on the model's real‐time performance, balancing learning efficiency in the early training stages with stability in the later stages. Furthermore, a novel reward function that considers distance approximation, heading alignment, and obstacle avoidance warning is constructed to alleviate the sparse reward problem. Experiments were conducted in an underwater simulation environment built using the Gazebo physics engine. The results demonstrate that, compared to DQN‐series algorithms, RFAD3QN achieves faster convergence during training and can generate smoother and shorter global paths in unknown environments. The simulation results validate the superiority of the proposed algorithm.
ABSTRACT Aiming at the mutual coupling of inherent interference, out‐of‐band leakage in the frequency domain and time‐domain energy spreading in the design of prototype filters for Filter Bank Multi‐Carrier (FBMC) systems, a time‐frequency collaborative optimization method for prototype filters based on the Constraint‐Aware Entropy‐Weighted Polar Lights Optimizer (CAEW‐PLO) is proposed. First, the independent coefficients of the prototype filter are taken as optimization variables, with three optimization objectives, harmful interference power, stopband energy in the frequency domain and overlapping energy in the time domain, constructed separately. A dynamic entropy‐weight mechanism is employed to adaptively adjust the contribution of each objective based on population target distribution. On this basis, in order to address the problems of insufficient initial coverage, weakened search directionality in the mid‐ to late stage and local stagnation in the high‐dimensional filter parameter space of the basic polar lights optimizer, chaotic reverse initialization, elite differential guidance and a stagnation‐triggered Cauchy mutation mechanism are introduced. Build an FBMC/OQAM simulation platform and conduct multiple independent comparisons with algorithms such as PLO, AO, ECO, HGS, HHO, SMA, and WOA under a unified function evaluation budget. The results show that CAEW‐PLO exhibits good comprehensive competitiveness in convergence performance, operational stability, and multi‐objective coordination ability. Further ablation experiments validated the effectiveness of various improvement mechanisms and three optimization objectives; system‐level experiments under communication conditions such as AWGN, Rayleigh, Rician, multipath and carrier frequency offset have shown that optimizing filters can reduce intrinsic interference and spectral leakage and achieve relatively balanced performance in terms of bit error rate, peak‐to‐average power ratio, spectral efficiency, and carrier frequency offset robustness.
ABSTRACT This paper proposes soft actor–critic with loop‐aware detection and event‐triggered stochastic escape. This is a reinforcement‐learning navigation algorithm for mobile robots operating in complex, unknown environments. The algorithm builds on soft actor–critic (SAC) and adds two cooperative modules. First, the real‐time loop detector can spot local optima traps during operation. It does this by checking the similarity of the robot's path over time and space, and by monitoring how often the robot changes direction in straight‐line speed. Frequent changes can cause the robot to move back and forth. When loops are detected, dynamic penalties guide the robot out of these traps. Second, the event‐triggered stochastic escape (ETSE) mechanism solves deadlocks in complex terrain. When the algorithm detects repetitive back‐and‐forth motions, the ETSE takes control. It injects high‐amplitude stochastic perturbations into the control signals to disrupt the deadlocked dynamics. This physical intervention physically forces the robot out of local minima and helps when normal reward signals do not guide learning. In Gazebo simulations, our method outperforms SAC, TD3, and DDPG in static and dynamic scenarios. Results show it adapts well and is robust in complex environments, providing a new lightweight option for autonomous robot navigation.
