Neco-v/MAC_scheduling

5G NR ISAC DRL-assisted MAC Scheduling Dataset

Overview

This dataset was generated as part of the 6G-SENSES project. It contains simulation traces that capture the network state and resulting performance metrics (rewards) in an environment that integrates communications and sensing (ISAC).

The goal of this dataset is to facilitate research and development of RL agents and other ML techniques for intelligent decision-making in the MAC scheduler of 5G networks.

The data comes from simulations performed in the ns-3 5g-LENA simulator, using the implementation available in the RBIS repository. This implementation includes a drift-plus-penalty scheduler for ns-3 5G-LENA responsible for the RB allocation among UEs, and an emulated xApp in which a DRL agent periodically adjusts the drift-plus-penalty weights to enhance performance and adaptability. For additional details about the scheduler implementation and its operation, readers are referred to this repository.

Simulation Context

The scenario simulates a single cell under realistic channel conditions, incorporating fast fading and 3GPP propagation models in an urban environment with buildings (UMa_Buildings scenario).

Specifically, the dataset was generated using the following configuration:

• —Network setup: 20 MHz bandwidth, 5G NR numerology 0, and 6 GHz carrier frequency.
• —User Equipments (UEs): 4 active UEs in total.
• —2 Virtual Reality (VR) UEs.
• —2 Cloud Gaming (CG) UEs.
• —Mobility: UEs are moving at a constant speed of 2 m/s using a Random Walk 2D mobility model (ns3::RandomWalk2dMobilityModel).
• —Traffic Models: 3GPP-compliant traffic models are used to generate application-layer traffic.
• —VR UEs: 10 Mbps data rate, 60 FPS, with a Guaranteed Flow Bit Rate (GFBR) of 2 Mbps.
• —CG UEs: 10 Mbps data rate, 60 FPS, with a GFBR of 4 Mbps.

Dataset Structure

The dataset contains tabular information where each row represents a simulation step (timestep) and contains the full RL environment transition: current state, action taken, reward obtained, and the next state.

The CSV columns are divided into the following categories:

1. Current State

The state of the network before taking the action. For each of the 4 UEs (i from 0 to 3), the following variables are categorized into 3 levels for simplicity:

• —mcs_i: Modulation and Coding Scheme (MCS).
• —q_i: Transmission buffer/queue occupancy level.
• —g_i: Virtual queue occupancy level.
• —obs_i: Binary indicator of obstacle presence that is expected to degrade the signal quality of UE i in the near future.

2. Throughput metrics

• —thput_i: Throughput obtained by the UE.
• —gfbr_i: GFBR required by the UE.

3. Actions

The decisions taken by the agent to adjust the scheduler weights.

• —action_v: Weight assigned to the reduction of Resource Block (RB) usage.
• —action_wq_i: Weight assigned to the queue of UE i.
• —action_wg_i: Weight assigned to the virtual queue of UE i.

4. Rewards

Feedback signals received after applying the action.

• —reward: Total aggregated reward calculated by the environment.
• —reward_thput: Reward component based on the obtained throughput.
• —reward_rbs: Reward component based on the assigned RBs.

5. Next State

The resulting state after applying the action. It follows the same structure as the current state:

• —next_mcs_i, next_q_i, next_g_i, next_obs_i (for i from 0 to 3).