Contents Machine Learning for Wireless Communications
Machine learning (ML) is drawing extensive attention as a result of recent advances in computing technologies, and its successful applications to image processing and natural language processing [1]. ML can also potentially increase the functionality and efficiency of radio networks [2]. In general, there are three popular classes of ML techniques:
Supervised learning infers functions from a labeled training dataset. It requires a supervisor, which trains the system parameters against the expected output of each input. Once trained, the system is able to make predictions with wide applications of classification, fault detection, and channel coding and decoding [3]. It is often used to approximate the function between the given input and output. Examples of supervised learning are neural network (NN), linear regression, and support vector machine (SVM). However, many supervised learning techniques may not suit wireless networks, because the learning accuracy highly depends on the representativeness of the training data. In addition, the performance extracted from supervised learning results is restricted by the representativeness of the training data.
Unsupervised learning infers functions to describe hidden structures from unlabeled data. It does not have a supervisor, and the system trains itself based on unlabeled input data. Examples of unsupervised learning are clustering algorithms, combinatorial algorithms, and game theory. It is often used to classify the input data into different classes based on the distribution of the data, and has been applied in resource optimization, mobility management, and smart caching in wireless cellular networks [2]. An example of unsupervised learning is adaptive resonance theory (ART), which is a clustering algorithm balancing between adaptability and stability at the same time. The learning process of ART can be influenced by different factors, such as vigilance value and input sequence. The performance of ART can deteriorate in constantly changing wireless communication networks, which requires quick and accurate response.
Reinforcement learning (RL) continuously labels real-time changing data in such a way that the learning framework is able to continuously evolve, adapting to the data. RL also produces rewards to incentivize the evolution. Take the popular Q-learning as an example. Q-learning does not rely on any a priori knowledge on the data and is able to learn from the data in an automated fashion. However, Q-learning undergoes the “curse of dimensionality” as the consequence of exponentially growing dimensions of the state-action spaces. The difference of RL from supervised and unsupervised learning is that RL operates based on a feedback mechanism where a reward is obtained if a “good” decision is made, or a penalty is given otherwise. By this means, the RL system can continuously evolve, while supervised learning and unsupervised learning typically provide static solutions. Q-learning has been applied to radio parameter configuration, coverage and capacity maximization, and resource optimization [4].
Other ML techniques include Markov model, heuristics, fuzzy controllers, and genetic algorithms [5]. These techniques are not part of the popular artificial intelligence algorithms. Nevertheless, the techniques can be easier to implement in many applications than supervised learning, unsupervised learning, and RL.
The authors discuss a different class of ML technique, stochastic online learning, and its promising applications to MEC. Based on stochastic gradient descent, stochastic online learning learns from the changes of dynamic systems rather than training data, decou-pies tasks between time sIoTs and edge devices, and asymptotically minimizes the time-averaged operational cost of MEC in a fully distributed fashion with the increase of the learning time.
