Can Deep Reinforcement Learning Revolutionize Time-Series Data Compression?

Introduction

With the proliferation of mobile Internet, IoT, and 5G, we have entered the era of digital economy where massive time‑series data is generated continuously. Efficient compression and storage of this data is a fundamental problem, and deep reinforcement learning offers a promising way to improve compression performance.

Background

Time‑Series Data

Time‑series data consists of timestamp‑value pairs and appears in many domains such as electrocardiograms, stock indices, and transaction logs. A time‑series database must handle massive queries, analysis, and predictions while also supporting high‑throughput reads, writes, and compression at the storage layer.

Typical representation uses two 8‑byte values for each point, leading to huge storage requirements that motivate advanced compression techniques.

Reinforcement Learning Overview

Reinforcement learning (RL) learns by interacting with an environment without requiring ground‑truth labels. Core elements are State, Action, and Reward. Common RL algorithms include Deep Q‑Network (DQN), Policy Gradient, and Actor‑Critic, each with different suitability for discrete or continuous action spaces.

Existing Time‑Series Compression Methods

Snappy – uses long‑distance prediction and run‑length encoding; widely used in InfluxDB.

Simple8b – delta encoding followed by packing based on a 16‑entry code table; also used in InfluxDB.

Compression Planner – combines generic tools (scale, delta, dictionary, Huffman, run‑length, patched constant) with static or dynamic selection.

ModelarDB – lossy compression based on user‑defined error tolerance, detecting linear patterns.

Sprintz – optimized for 8/16‑bit integers using scaling, delta encoding, and bit‑level packing.

Gorilla – lossless compression used at Facebook, employing delta‑of‑delta for timestamps and XOR‑based transformation with Huffman coding.

MO – similar to Gorilla but removes bit‑packing for byte‑aligned operations, trading compression ratio for speed.

Characteristics of Large‑Scale Time‑Series Compression

Time correlation – strong temporal locality with regular sampling intervals.

Pattern diversity – wide variety of signal shapes and precision requirements.

Data massiveness – daily workloads can reach petabyte scale, demanding high‑throughput algorithms.

Two‑Stage Deep Learning Compression Framework (AMMMO)

The compression process is divided into a Transform stage that maps raw data to a more regular space, followed by a Differential Coding stage that efficiently encodes the transformed differences.

Six basic transform primitives and three differential coding primitives are defined, which can be combined to form many compression modes. Direct exhaustive combination is impractical due to high control‑information overhead.

AMMMO (Adaptive Multiple‑Mode Middle‑Out) first determines global timeline characteristics (nine control parameters) and then selects the best compression mode from a small candidate set (≈4 modes) for each block, reducing control overhead to about 3%.

Rule‑Based Mode Selection

A scoring system (scoreboard) evaluates each mode across the timeline, selecting the optimal one based on statistical analysis. This approach, however, requires manual tuning and extensive code maintenance.

Deep Reinforcement Learning for Mode Selection

The mode‑selection problem is modeled as a multi‑label classification task. A fully‑connected MLP processes blocks of 32 points (≈256 B) and outputs probabilities for each control parameter via a region‑softmax.

During training, M blocks are sampled, each replicated N times, and fed through the network. The resulting parameters are applied to the underlying compressor, and the compression ratio is used to compute a custom loss that combines a reward term (fn) with cross‑entropy regularization (Hcs, H).

Policy Gradient is chosen over DQN and Actor‑Critic because the action space is discrete but not smoothly ordered, and the network complexity is modest.

Experimental Evaluation

Design

Datasets include 28 large timelines from Alibaba Cloud IoT and server workloads, as well as the UCR time‑series benchmark. Baselines are Gorilla, MO, and Snappy. AMMMO is evaluated with four parameter‑selection strategies: Lazy, rnd1000Avg, Analyze (hand‑crafted), and ML (deep RL).

Compression Ratio

AMMMO achieves roughly 50% higher compression ratios compared to Gorilla/MO across all test sets. The ML‑driven selection slightly outperforms the hand‑crafted Analyze method and far exceeds random averaging.

Runtime Performance

Implemented on an Intel CPU 8163 and Nvidia P100 GPU, AMMMO attains GB/s throughput for both compression and decompression, comparable to MO while offering superior compression.

Learned Insights

Parameter heatmaps reveal that the network learns meaningful patterns, such as consistent byte‑level operations (delta, XOR) and heightened activity on the most variable bytes, indicating that the model captures genuine compression heuristics.

Conclusion

Deep reinforcement learning can effectively automate the selection of compression modes for massive time‑series data, delivering significant storage savings and high‑speed processing while providing interpretable learned behaviors.