NTT DATA 5G Latency Optimization RL Prediction Model (ONNX)

Model Info This repository contains a machine learning model developed by NTT DATA as part of the MLSysOps project. This artifact provides an ONNX export (opset 18) of a Deep Q-Network (DQN) agent trained to select the best data center (among 3) for a client request in a 5G/MEC setting, optimizing latency-related outcomes (and incorporating carbon intensity as a feature).Given the current state of three candidate data centers, the model outputs Q-values for each possible selection and chooses the data center with the highest Q-value. Purpose This model performs a Reinforcement Learning agent that predict a discrete value: Input: A normalized tensor representing the current state of three candidate data centers, shaped as (3 × 10), where each row corresponds to a data center and each column represents a feature such as client identifier, resource utilization, network metrics, latency statistics, packet loss, and carbon intensity. The input is MinMax-scaled using parameters learned during training and includes a label-encoded client identifier. Output: A discrete action in {0, 1, 2} corresponding to the selection of the optimal data center among the three candidates, computed as the index with the highest predicted Q-value. Repository Structure A typical layout (as used in the accompanying repository/bundle): model/ 5g_latency_opt_dqn_model.onnx — ONNX model (DQN Q-network) model_config.json — model metadata, I/O specs, feature order, preprocessing parameters src/ state_serializer.py — builds the model input tensor from JSON scenarios and applies preprocessing minmax_scaler.py — lightweight MinMax scaling implementation (training-fitted parameters) inference_engine.py — ONNXRuntime inference wrapper (argmax over Q-values) action_interpreter.py — converts predicted action to a human-readable decision demo.py — end-to-end demo (JSON → preprocess → ONNX inference → decoded action) requirements.txt — minimal Python dependencies Training Data The model was trained on a tabular dataset containing per-data-center telemetry and network metrics. Each decision step groups 3 rows (one per candidate data center) into a single observation. Features Used The training preprocessing: MinMax normalization for numeric features: cpu_usage_percent memory_usage_percent disk_usage_percent net_in_percent net_out_percent latency_avg latency_mdev lost_percent carbon_intensity Label encoding for: client_id Dropped columns: start_time, end_time, net_in_absolute, net_out_absolute, latency_min, latency_max Important: the inference pipeline must reuse the same MinMaxScaler parameters (data_min/data_max) and the same client_id encoding mapping fitted during training. Model Architecture The exported ONNX model contains the SB3 DQN policy Q-network (MLP-based Q-function). The network maps a (3×10) observation (three candidate data centers, ten features each) to Q-values for the three discrete actions (select DC0/DC1/DC2).At inference time, the recommended decision is argmax(Q-values). Model Specification Inputs Name: observation Type: float32 Shape: (batch_size, 3, 10)where: dimension 1 = candidate data centers (always 3) dimension 2 = feature vector per data center Feature Order (Last Dimension) The last dimension (size 10) follows this exact order: client_id (label-encoded) cpu_usage_percent (MinMax-scaled) memory_usage_percent (MinMax-scaled) disk_usage_percent (MinMax-scaled) net_in_percent (MinMax-scaled) net_out_percent (MinMax-scaled) latency_avg (MinMax-scaled) latency_mdev (MinMax-scaled) lost_percent (MinMax-scaled) carbon_intensity (MinMax-scaled) Outputs Name: q_values Type: float32 Shape: (batch_size, 3) Meaning: Q-values for the three actions: action 0 → select Data Center 0 (Milan) action 1 → select Data Center 1 (Rome) action 2 → select Data Center 2 (Cosenza) Limitations The model is trained for exactly 3 candidate data centers; input shape is fixed to (3,10). Correct behavior depends on identical preprocessing: MinMax scaling must use training-fitted min/max values client_id must be encoded using the training-fitted mapping (unknown IDs should be handled explicitly) Generalization outside the training distribution (different telemetry ranges, unseen client populations, different operational conditions) is not guaranteed. This model provides a decision policy but does not guarantee optimality; it should be validated in the target deployment setting before use. Usage Demo Setup Environment Create a Python environment and install dependencies: python -m venv .venv source .venv/bin/activate # (Linux/macOS) # .venv\Scripts\activate # (Windows) pip install -r requirements.txt At minimum, the runtime requires: onnxruntime numpy pandas Run Inference Script Run the end-to-end demo: python demo.py The demo will: Load a JSON scenario containing dataCenterStates (3 entries) Apply preprocessing (MinMax scaling + client_id encoding) Run ONNX inference via ONNXRuntime Print the chosen data center index (argmax over Q-values) Citation If you use this model/artifact in academic work, please cite it as: @software{pazienza_5g_latency_opt_dqn_onnx, title = {5G Latency Optimization DQN (ONNX)}, author = {Andrea Pazienza, Massimiliano Rossi}, year = {2026}, publisher = {Zenodo}, doi = {10.5281/zenodo.18303750}, url = {https://doi.org/10.5281/zenodo.18303750} } Acknowledgement & Funding This work is part of the MLSysOps project, funded by the European Union’s Horizon Europe research and innovation programme under grant agreement No. 101092912. More information about the project is available at https://mlsysops.eu/