Multi-Batch RL Distillation for Base Station Traffic Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Reinforcement learning algorithms face challenges in real-world applications due to poor data efficiency and inflexibility in offline learning, particularly in scenarios with limited interaction samples, limiting their effectiveness in tasks like traffic load balancing in communication systems.

Innovation Solution

A method involving multi-batch reinforcement learning using imitation learning and knowledge distillation to create a generalized AI model for predicting future traffic loads across multiple base stations, leveraging data augmentation and transfer learning to improve policy generalization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If batch RL algorithms are designed for a single batch setting with large fixed data sets, then learning reliability is improved, but adaptability to limited data scenarios deteriorates

Engineering Contradiction:
Improvelearning reliabilityVSAvoidadaptability to limited data
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent creates a unified batch RL framework that can handle both single-batch and multi-batch scenarios through a common architecture. The algorithm universally processes batches by augmenting state-action pairs from multiple sources (original batch, simulated data, other batches) through a consistent data augmentation pipeline, making it adaptable to various data availability conditions while maintaining learning reliability.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent segments the learning process into distinct phases: data augmentation phase (generating synthetic state-action pairs), policy learning phase (training the control policy), and generalization phase (applying to unseen batches). This segmentation allows the algorithm to systematically handle limited data by combining multiple data sources in controlled stages, improving both reliability and adaptability.

Inventive Principle:
Principle #1Segmentation

2Reliability

If online RL algorithms are used to learn optimal control policies, then performance on control tasks is improved, but data efficiency deteriorates due to large number of interactions required

Engineering Contradiction:
Improvecontrol policy performanceVSAvoiddata efficiency
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent performs preliminary data augmentation by generating synthetic state-action pairs from simulated environments and other batches before the actual policy learning occurs. This preliminary preparation of augmented data reduces the need for extensive real-world interactions during the learning phase, improving data efficiency while maintaining control policy performance through pre-prepared training data.

Inventive Principle:
Principle #10Preliminary action

3Quantity of substance

If imitation learning is applied to learn control policies from demonstrated data, then sample efficiency is improved, but policy generalization to unseen scenarios deteriorates

Engineering Contradiction:
Improvesample efficiencyVSAvoidpolicy generalization
Core Design Contradiction:
Quantity of substanceVSAdaptability or versatility

Solution Approach 1:

The patent adds a temporal dimension to imitation learning by incorporating data from multiple batches collected at different times and conditions. Instead of learning from a single static dataset, the algorithm processes sequences of batches, augmenting state-action pairs across time dimensions. This multi-temporal approach improves policy generalization to unseen scenarios while maintaining sample efficiency through the imitation learning framework.

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

Data Source

PatentEP4324241B1Method and system for multi-batch reinforcement learning via multi-imitation learning
Publication Date: 2026.01.28 SAMSUNG ELECTRONICS CO LTD
  • EP4324241B1 patent drawingFigure 1~2
  • EP4324241B1 patent drawingFigure 3
  • EP4324241B1 patent drawingFigure 4~6

AI summary

A server may receive a first traffic data and a second traffic data from a first base station and a second base station; obtain a first augmented traffic data for the first base station, based on the first traffic data and a subset data of the second traffic data; obtain a second augmented traffic data for the second base station, based on the second traffic data and a subset data of the first traffic data; obtain a first artificial intelligence (AI) model via imitation learning based on the first augmented traffic data; obtain a second AI model imitation learning based on the second augmented traffic data; obtain a generalized AI model via knowledge distillation from the first AI model and the second AI model; and predict a future traffic load of each of the first base station and the second base station based on the generalized AI model.