AI Learns to Keep Streaming Video Hot and High Quality

Every time you press play on a video, a hidden negotiation takes place inside a data center. The streaming service must decide which version of the clip to serve — a crisp high-bitrate file or a leaner, lower-quality copy — while the storage system must decide where that file lives: on fast, always-on disks that burn power around the clock, or on slow, energy-sipping drives that can be spun down when nobody is watching. A new study published in Cluster Computing shows that a deep reinforcement learning agent can learn to make these decisions far better than the rule-based systems that storage engineers have relied on for years, improving delivered video quality by 1.9% to 31.4%, with an average gain of 10.9% across different hot storage capacities.

The research, carried out by Kyeongmin Kim and Minseok Song of Inha University in Korea together with Younghyun Kim of Purdue University in the United States, tackles a problem that has quietly grown enormous as video dominates internet traffic. Modern video-on-demand platforms do not store a single copy of each clip. To support adaptive bitrate streaming — the technology standardized as Dynamic Adaptive Streaming over HTTP, or DASH — every video is encoded in multiple bitrate versions, and often split into segments, so that playback can shift smoothly between quality levels as network conditions fluctuate. Multiply that by millions of videos and their wildly uneven popularity, and efficient storage management becomes a genuinely hard combinatorial puzzle.

Traditional approaches lean on a simple dichotomy: hot data and cold data. Files that are accessed frequently are classified as hot and kept on active disks ready to serve requests instantly. Infrequently accessed files are designated as cold and migrated to deactivated disks, which are powered down to save energy. This hot-cold classification has historically depended on historical access patterns — count the requests over a window, sort the files, and draw a line. The strategy works reasonably well when popularity is stable, but the authors argue it fails to adapt to dynamic, content-aware factors such as shifting popularity or changing quality demands. A video that was obscure yesterday can explode in views overnight, and a storage system that only looks at request counts reacts too slowly, wasting scarce hot storage on content nobody wants at high quality while starving the clips that viewers are actively streaming.

The energy stakes are considerable. Data centers already consume a substantial and rising share of global electricity, and hard drives that remain active solely to serve occasional requests represent a classic inefficiency. Powering down idle disks is one of the most effective energy-saving techniques in storage systems, which is precisely why hot-cold tiering became standard practice. But every gigabyte of hot storage is expensive in watts, so the question is not merely which files are hot, but which bitrate versions of which segments deserve to occupy the limited hot zone. Serving a popular video from a low-quality version when a high-quality copy could have been stored instead directly degrades the viewer’s experience.

Kim and colleagues reframe this allocation problem as a sequential decision-making task and hand it to deep reinforcement learning, the family of techniques behind some of the most celebrated results in modern artificial intelligence. In a reinforcement learning setup, an agent interacts with an environment, observes its state, takes actions, and receives rewards; over many iterations it learns a policy that maximizes cumulative reward. Deep reinforcement learning adds neural networks to approximate that policy, allowing the agent to generalize across states it has never explicitly encountered. The framework the researchers built constructs an action space for bitrate version selection — the agent decides which quality version of each video segment to place in hot storage — and an observation space that incorporates content-aware features including popularity, a quality index, and storage size. A reward model directs the agent toward quality-optimized decisions, so the learning signal is aligned with the actual goal: sustaining high service quality under unpredictable changes in request patterns and network conditions.

The choice of observation features is what distinguishes this work from earlier DRL-based storage systems that focused mainly on power consumption or access frequency. By feeding the agent popularity dynamics and quality-related indices, the researchers give it the raw material to reason about content-aware trade-offs: a segment of a surging viral video with high quality demand is worth more hot storage than a segment of a fading clip, even if their raw request counts look similar. The agent learns these relationships from experience rather than from hand-tuned thresholds, which is exactly the adaptability that static classification schemes lack. The authors trained their agent using proximal policy optimization, a widely used and stable policy-gradient algorithm, with implementation details drawn from established open-source baselines.

Reinforcement learning alone, however, cannot guarantee that the resulting allocations fit within physical capacity limits. The researchers therefore complemented the learned policy with a greedy algorithm that allocates hot storage space for each video segment under capacity constraints. The combination is elegant: the DRL agent supplies quality-oriented judgments about which bitrate versions matter, while the greedy allocator — a technique familiar from classical knapsack problems in combinatorial optimization — turns those judgments into a feasible placement plan that respects the finite size of the hot zone. This two-stage design keeps the system practical, because a learned policy that routinely proposes infeasible allocations would be useless in production storage.

The experimental results are striking in their consistency. Across different hot storage capacities, the proposed scheme improved video quality by 1.9% to 31.4% compared with standard benchmarks, averaging a 10.9% gain. The improvement is measured in perceptual quality terms, consistent with the video quality metrics that major streaming platforms use to validate viewer experience. The range itself is informative: when hot storage is plentiful, the margin over conventional schemes narrows because even imperfect heuristics can store most high-value versions; when hot storage is scarce, the learned policy’s ability to discriminate between segments pays off dramatically, which is where the largest gains appear.

For the storage industry, the implications reach beyond video-on-demand. Hot-cold classification underpins tiering in cloud block storage, archival systems, and content delivery networks, and DRL has already been applied to cache replacement, metadata management, and adaptive replication in distributed storage. What this study adds is a template for injecting content-aware signals — popularity trajectories, quality indices, encoding sizes — into the learning loop, so that storage decisions track what viewers actually experience rather than what request logs record. As platforms grapple with multi-version libraries and fluctuating demand, that distinction could determine whether scarce, power-hungry hot storage is spent on the content that matters.

There are caveats, as with any learning-based system. Reinforcement learning agents require training time and can be sensitive to the distribution of request patterns they encounter, and the authors note that no datasets were generated or analysed during the current study, meaning evaluation relied on modeled workloads rather than production traces. Real deployments would need mechanisms to retrain or fine-tune the policy as content catalogs and viewer behavior evolve. Still, the reported gains — an average 10.9% quality improvement from smarter placement of the same bytes — come essentially for free in computational terms, at a time when every percentage point of data center energy efficiency matters. If the next generation of storage systems can learn to keep the right data hot, viewers may get better streaming without the planet paying for it in watts.

Subject of Research: Computer Science

Article Title: Quality-oriented hot data management using deep reinforcement learning in video storage systems

Article References: Kim, K., Kim, Y., & Song, M. (2026). Quality-oriented hot data management using deep reinforcement learning in video storage systems. Cluster Computing, 29(15), Article 853. https://doi.org/10.1007/s10586-026-06681-x

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06681-x

Keywords: Quality-oriented, data, management, deep, reinforcement, learning, video, storage, systems, scientific research