Introduction
As we transition from 5G to 6G, wireless networks face unprecedented challenges. Future networks must support up to 10 million devices per square kilometer and facilitate complex machine-to-machine (intelligence-to-intelligence) communications. Current "adaptive" technologies, such as Adaptive Modulation and Coding (AMC), are increasingly insufficient. These systems rely on pre-defined, static policy sets and centralized intelligence, making them too rigid to handle the massive data and diverse scenario-semantic requirements of 6G. To overcome these limitations, a shift from mere adaptation to autonomous "self-evolution" is required.
The Self-Evolving Framework
The proposed framework moves beyond choosing from fixed databases to generating entirely new communication strategies. It is organized into three hierarchical layers that facilitate a continuous cycle of learning and optimization:
- Data Layer: The physical foundation comprising transceivers and channels responsible for raw signal transmission.
- Information Layer: This layer senses the environment, recognizes specific application scenarios, and makes intelligent decisions regarding waveforms and protocols.
- Knowledge Layer: Acting as the system's "brain," this layer generates, evaluates, and reconstructs a knowledge base.
At the heart of this architecture is the Knowledge Chain. Insights gathered at the information layer are transformed into "knowledge" that is fed back to optimize the lower layers. This allows the system to update its "communication genome"—its source codes, modulation schemes, and protocols—based on real-world environmental feedback.
Algorithmic Innovations and Performance
To validate this self-evolving concept, the researchers introduced two primary algorithms:
- SaE-ELM (Self-adaptive Evolutionary Extreme Learning Machine): This algorithm utilizes differential evolution to optimize the internal parameters of a neural network for signal detection. In simulations under Rayleigh fading conditions, SaE-ELM demonstrated superior reliability, outperforming standard Extreme Learning Machines by approximately 2 dB in Bit Error Rate (BER) and showing a 6 dB gain over traditional non-ML baselines like Zero-forcing.
- SE-QL (Self-Evolving Q-Learning): Designed for complex coordination tasks like UAV rendezvous in space and frequency, this algorithm combines reinforcement learning with genetic algorithms. Results showed that SE-QL reached convergence 80% faster than standard Q-learning, proving that evolutionary memory significantly enhances learning efficiency in multi-agent scenarios.
Future Directions and Challenges
While promising, several hurdles remain for the realization of self-evolving 6G networks:
- Environmental Complexity: Current models struggle to characterize the intense, dynamic interference of 6G environments across the "six domains."
- Computational Overhead: The evolutionary process of mutating and testing policies is resource-intensive. Solutions like Mobile Edge Computing (MEC) and Federated Learning will be critical to distributing this load.
- Security: As networks become more intelligent, they also become targets for "self-evolving" attackers. Real-time, learning-based attack detectors will be necessary.
- Knowledge Organization: Constructing a robust network "brain" will likely require the integration of Digital Twins and Knowledge Graphs to effectively organize and utilize evolved insights.
Summary updated