<![CDATA[The proliferation of computationally intensive Industrial Internet of Things (IIoT) applications requires offloading policies that reduce latency and energy consumption under constrained device, edge, and cloud resources. Although reinforcement learning (RL) is suitable for sequential offloading decisions, its performance is highly sensitive to hyperparameters that are often fixed or manually tuned. The research contribution is an automated QLSO framework that uses the Snake Optimizer (SO) to tune the learning rate, discount factor, and exploration-decay rate of tabular Q-learning (QL) in a three-tier IIoT-edge-cloud architecture. Unlike generic external Hyperparameter Optimization (HPO) methods, QLSO embeds SO as a derivative-free, population-based tuning layer within the QL offloading workflow, where candidate hyperparameters are evaluated based on latency-energy-aware performance under three-tier IIoT resource constraints. The problem is formulated as a Markov decision process with discrete local, edge, and cloud execution actions, and candidate hyperparameters are evaluated through episodic training followed by greedy validation. Across five independent random seeds, QLSO converged about 1.5× faster in the complex scenario and achieved lower latency and energy than the QL and SARSA baselines, while remaining close to the strongest raw-metric baseline, DQN (0.0493 s and 1.72×10⁻² J for QLSO versus 0.0488 s and 1.65×10⁻² J for DQN). However, QLSO achieved the highest latency-optimal decision rate (49.0%, +3.3 percentage points over DQN), and the ablation study showed a 9.1% fitness improvement over fixed-parameter QL. The framework remains simulation-based and single-agent, requiring validation with partial offloading, dynamic network models, and realistic platforms such as NS-3, CloudSim, EdgeCloudSim, or physical IIoT testbeds.]]>
