The 125th Installment
Mission of AI for Next-generation Communication

by Zhang Chaofeng,
Assistant Professor

IoT and the new revolution it will bring

At present, the penetration rate of IoT is an important metric for an information society. In particular, the importance of the spread of medical IoT devices with COVID-19 has increased significantly. Terminals used for IoT usually have fixed standards for computing, communication, and caching. IoT plays an important role in interactions between humans and computers. Therefore, a stable and compatible IoT control framework is one challenge for the future.

3C （Communication，Computation and Caching）

The concept of 3C has also been mentioned in a growing number of IoT studies, specifically communication, computing, and caching. Communication refers to device-to-device (D2D) and device-to-base station (D2B) communication. With the spread of 5G, the issue of spectrum utilization has also become important metric for the IoT framework. Computing refers to the computing power of IoT terminals. With the spread of AI, the number of devices that require preprocessing of acquired information, such as facial recognition and behavior recognition, is increasing. Some computing power is required to pre-process light computational tasks. After processing, the compressed data containing the characteristics of the raw data is uploaded to a server. Caching refers to the overall accessible cache of the framework, not individual equipment. For example, in a content centric network (CCN), only popular data is stored in a local area and distributed over short distances as needed, which reduces bandwidth usage. The metrics, communication, calculations, and caching of IoT systems, as well as more efficient coordination, can improve service availability and user experience for next-generation networks.

Tradition-based Innovation

Traditional network evaluation metrics are primarily based on a single metric. For example, in 4G communications, optimization goals such as throughput, delay response, arrival rate, and so on are common. In particular, routing algorithms based on data analysis usually predict future trends through the current state of cognitive networks and spectrum resources. Furthermore, the computing resource allocation problems faced by distributed systems need to be further optimized by heuristic algorithms depending on the device and communication conditions. At the same time, essential research to improve the quality of experience (QoE) also includes metrics such as capacity issues, content popularity, and so on. For a variety of metrics, the traditional single goal optimization approach becomes a trade-off issue (trading between each goal). For example, excess bandwidth is occupied to improve the performance of a distributed AI learning system. However, the likelihood of blocking communication channels is also higher. Also, to increase computational efficiency, large amounts of raw data are pre-stored, consuming large storage space while increasing energy consumption. In the face of 3C diversification metrics and to make more efficient use of the resources of individual devices, the ability to control the overall framework is needed.

AI is an essential technology for next-generation communication

The deep reinforcement learning (DRL) scheduling algorithm is a useful solution for next-generation network resource management. This algorithm operates to acquire the current network state and manages 3C resources. When processing a task, the terminal determines the range of available facilities based on the surrounding conditions to be monitored. AI learning means that for a given network state, reward scores are listed with various assignments (e.g., reward scores for energy consumption reduction and task delay). DRL can provide a list of Q-values from each network state. Each Q-value reflects the reward of the corresponding hypothetical action. Of course, the best strategy for non-human computers is to take actions with the maximum Q-value calculated by themselves. The learning process can be done offline by distributing trained neural networks to each terminal.

Conclusion

This solution is a way of adapting human thinking to solve problems that are inherently complex. There is the Chinese saying, “To have eyes but fail to recognize Mt. Tai,” which means to be blind to something great. As a weakness, neural networks, like humans, usually make optimal choices only based on surrounding interests and short-term interests. They make assessments based on a variety of unique criteria and choose actions with higher scores.

Neural networks can certainly solve the problem of increasing data dimensions, but with the growth of terminals and the cloud, research on integrated mechanisms, such as federal learning, must go further.