Divide-and-Conquer Decomposition Strategies for Real-Time Artificial Intelligence Decision Systems: A Comparative Survey

Abstract

Abstract—Real-time artificial intelligence (AI) decision systems must resolve complex inputs — a video frame, a network flow, a multi-agent traffic scene — within a bounded time budget. A recurring, if rarely named, strategy for meeting this budget is the classical divide-and-conquer (D&C) paradigm: partition the problem into smaller, largely independent subproblems, solve each subproblem locally, and combine the partial results into a final decision. This survey makes that pattern explicit. We first formalize the D&C template (divide, solve, combine) and relate it to real-time scheduling theory, RTOS design, and kernel-level execution protection, which together bound the time cost of each stage. We then read nine deployed or proposed real-time AI systems — spanning modular inference pipelines, classification tasks, graph matching, transportation, energy markets, trajectory prediction, bot detection, federated intrusion detection, and honeypot-based defense — through this lens, and build a taxonomy of how each system realizes the divide step, the solve step, and the combine step under its own timing constraints. We further discuss adversarial robustness and low-code/no-code governance as cross-cutting concerns for decomposed pipelines. We conclude that most deployed real-time AI decision systems already implement an implicit divide-and-conquer structure, and that making this structure explicit clarifies where scheduling guarantees, security boundaries, and combine-step errors each need to be independently verified. Index Terms—Divide-and-conquer, real-time systems, artificial intelligence, decomposition, distributed decision-making, federated learning, scheduling.