The main objective of PORTRAIT project is to carry out the proof of concept of a deep reinforcement learning solution for RAN slicing in an industrially relevant environment, paving the way for market exploitation and technology transfer.

The project general objective will be achieved through a set of specific objectives:


Objective #1: To specify the PoC platform

Sub-objective #1.1:

To define the PoC architecture and functional model. This will take as the basis the current architecture of the 5GCAT pilot and will determine the components for integrating the DQN-MARL RAN slicing solution on the management and orchestration platform and the interactions with the RAN nodes through the O-RAN interfaces.

Sub-objective #1.2:

To define the validation and test plans. This will specify the different tests to be performed to ensure the correct operation of the DQN-MARL RAN slicing solution integrated in the PoC platform and the appropriate communication with the rest of involved modules. This will also specify the tests and metrics to be conducted in order to demonstrate the behavior and performance of the proposed solution in the 5GCAT use case.

Objective #2: To develop the software for implementing the DQN-MARL RAN slicing solution

Sub-objective #2.1:

Development of the trainer. This component constitutes the training part of the DQN model intended to learn the neural network parameters that determine the per-slice action selection policies to be used by the RAN slice manager. It includes a network simulator that mimics the behavior of the real network and that is fed by training data consisting of multiple time patterns of the required capacity of the slices in the different cells and several agents that learn the decision making policies.

Sub-objective #2.2:

Development of the RAN slice manager. This component includes the inference part of the DQN model and the functions needed to configure the amount of allocated resources per slice in the RAN nodes through the O1 interface. It makes use of the policies learnt by the trainer.

Sub-objective #2.3:

Development of the interfaces. This includes both the internal interface that will be used for the communication between the trainer and the RAN slice manager for transferring the learnt policies and the communication with the gNBs of the PoC platform through the O1 interface for collecting the performance measurements from the network and for transferring the capacity allocated per slice.

Sub-objective #2.4:

Development of the dashboard. This will be required to configure the input parameters of the DQN-MARL RAN slicing solution, including the hyperparameters of the DQN trainer and the parameters of the Network Slice Instance (NSI) profile that determine the SLA in terms of throughput per slice, throughput per user equipment (UE) and terminal density. The dashboard will also allow monitoring the operation of the solution by presenting different metrics and performance indicators (e.g. evolution of the capacity allocated per slice, evolution of the throughput per slice, etc.).

Objective #3: To conduct the PoC of the DQN-MARL RAN slicing solution

Sub-objective #3.1:

To integrate the DQN-MARL RAN slicing solution in the PoC platform. This will consist in integrating the developed trainer, RAN slicing manager and dashboard components of the solution in the platform of the 5GCAT pilot and on testing and validating their operation and the interactions among them and with the rest of components of the platform, according to the plans defined in Sub-objective #1.2.

Sub-objective #3.2:

To conduct the training of the DQN-MARL RAN slicing solution. This will make use of the developed trainer to learn the action-selection policies that determine the capacity allocated to each slice in each cell prior to executing the solution on the real network. The training process relies on training datasets created from real network measurements combined with synthetic data.

Sub-objective #3.3:

To carry out demonstrations and performance assessment. The demonstrations target the evaluation of the solution in terms of both practical aspects (e.g. execution speed, training speed) and performance aspects (e.g. Service Level Agreement satisfaction of the different slices, throughput, etc.), according to the plans defined in Sub-objective #1.2.

Objective #4: To define and execute an exploitation and technology transfer plan

The project will define and execute a plan for the industrial exploitation of the DQN-MARL RAN slicing solution. For this purpose, various forms of partnerships will be explored, from small cell vendors to venture builders, and a business plan will be created. Moreover, the project will disseminate its results in relevant fora (e.g. exhibitions and events, visits to stakeholders, etc.) in order to establish the appropriate contacts that facilitate these partnerships.

Objective #5: To foster entrepreneurial and innovation spirits in the research team

The research team sees PORTRAIT as an opportunity to expand the horizon of its current activities towards a tangible technology transfer to the industry, taking the MARL-DQN RAN slicing solution as the starting point for making this leap. In this way, all the development, testing, promotion and exploitation activities conducted within the project have the strategic objective of promoting the culture of technology transfer and innovation in the research team, with the target that this spirit materializes in new initiatives beyond the project lifetime.