Distributed Multi-Agent DRL for Cellular Handover Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current cellular network handover optimization methods face challenges in efficiently managing handovers in complex, distributed environments due to limitations in existing reinforcement learning algorithms, particularly in offline training and multi-agent settings, which affect Quality of Experience (QoE) and require centralized data aggregation, leading to latency and inefficiencies.

Innovation Solution

A distributed multi-agent Deep Reinforcement Learning (DRL) system is deployed on base transceiver stations, where computing agents communicate to make handover decisions using a trained policy based on signal strength and network state information, optimizing handovers in a fully offline manner to avoid disrupting network operations and improve scalability and fault tolerance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Power

If centralized cloud environment is used for handover optimization, then computational power is sufficient, but latency increases and distributed resources are underutilized

Engineering Contradiction:
Improvecomputational powerVSAvoidlatency
Core Design Contradiction:
PowerVSLoss of time

Solution Approach 1:

The patent segments the centralized handover optimization problem into multiple distributed sub-problems, with each base station running its own reinforcement learning agent. This segmentation allows local decision-making at the edge (base stations) while maintaining sufficient computational power through distributed processing, thereby reducing latency compared to centralized cloud processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from a single centralized optimization dimension to multiple distributed dimensions by deploying RL agents across numerous base stations. This dimensional shift enables parallel processing of handover decisions across different geographic locations,充分利用 distributed computational resources while minimizing latency through local real-time decision-making.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Productivity

If online reinforcement learning is used for handover optimization, then real-time learning is possible, but network operations are disrupted by random action testing

Engineering Contradiction:
Improvereal-time learning capabilityVSAvoidnetwork operation stability
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies preliminary action by pre-training reinforcement learning agents offline using historical network data before deployment. This preliminary training phase allows the system to learn optimal handover policies without disrupting live network operations. Once trained, the agents deploy their learned policies to real-time network operations, maintaining both productivity and reliability.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces an intermediary training environment that separates the learning process from the production network. The RL agents are trained in this intermediate offline environment using historical data, then their learned policies are transferred to the live network. This intermediary approach enables real-time learning capabilities while protecting network stability by preventing random exploration actions in production.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Adaptability or versatility

If multi-agent distributed system is deployed at base stations, then scalability and fault tolerance improve, but system complexity increases

Engineering Contradiction:
ImprovescalabilityVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements self-service by enabling each base station to autonomously run its own reinforcement learning agent for local handover optimization. Each agent independently processes local network conditions and makes decisions without requiring complex centralized coordination. This self-service approach enhances scalability and fault tolerance while managing system complexity through decentralized autonomy.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent manages multi-agent system complexity through parameter standardization and uniformity. All RL agents use the same algorithmic framework, state representations, and action spaces, allowing for consistent behavior across distributed base stations. This parameter uniformity enables scalable deployment of multiple agents while controlling overall system complexity through standardized configurations.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4391644A1Handover optimisation in a cellular network by multi-objective deep reinforcement learning applied by agents deployed in base-stations
Publication Date: 2024.06.26 TELEFONICA INNOVACION DIGITAL SL
  • EP4391644A1 patent drawingFigure 1
  • EP4391644A1 patent drawingFigure 2~3
  • EP4391644A1 patent drawingFigure 4

AI summary

A method, distributed system, device and computer programs for efficient and distributed network optimization of a cellular network are proposed. The cellular network is distributed in different cells where each cell comprises a base transceiver station providing network coverage to a plurality of computing devices, each computing device being associated to a user. The method comprises associating a computing agent to each base transceiver station; receiving, by a given computing agent, from at least one given computing device connected to its associated base transceiver station, signal strength information of the given computing device, the signal strength information being measured on the given computing device for the associated base transceiver station and for neighboring cells; determining, by the given computing agent, a handover decision for said given computing device by means of executing a reinforced learning decision-making algorithm on an established trained policy and on the received signal strength.