Online Confidence-Gated LSTM-DQN for Dynamic Cloud Resource Allocation
TL;DR
A causally consistent, confidence-gated LSTM-DQN scheduler that trains stably and safely on every seed, holds response time and SLA violations in line with standard RL and simple heuristics, and clearly outperforms a metaheuristic-augmented Q-learning baseline.
Abstract
Cloud schedulers that pair workload prediction with reinforcement learning (RL) rarely check whether a given prediction can actually be trusted, and earlier confidence-gated designs often mix current and future information inconsistently. We fix that inconsistency and build a causally consistent, confidence-gated LSTM-DQN scheduler: an LSTM forecasts next-step workload, a retrospective, error-based confidence score gates how much a Deep Q-Network (DQN) scheduler leans on that forecast, and only information available at decision time is ever used. We implement and pilot-test this architecture in a Python-based discrete-event simulation configured to match a CloudSim-style environment (10 hosts, 30 VMs), benchmarking it against FCFS, Round Robin, standard RL, two ablation variants, and two simplified state-of-the-art comparators across five random seeds. The results show the method works as intended: it trains stably and safely on every seed, holds response time and SLA violations in line with standard RL and simple heuristics, and clearly outperforms a metaheuristic-augmented Q-learning baseline, which suffered severe instability under the same conditions. Code, raw results, and statistical tests are released for independent verification, with scaled-up training identified as the natural next step to test whether larger performance gains emerge.