Power Amplifier Direct Modeling Using Machine Learning
Sumanen, Matias (2020)
2020
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2020-11-18
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202010257464
https://urn.fi/URN:NBN:fi:tuni-202010257464
Tiivistelmä
The power amplifier (PA) is an important component of a wireless transceiver. Power amplifiers work non-linearly, especially when good power efficiency is emphasized. They amplify the input signal but simultaneously, they also distort the signal. In this theses, the aim is to model power amplifiers using machine learning.
Power amplifier modeling is classically conducted using polynomials. It can also be done using machine learning. However, modeling them with machine learning is much more uncommon. According to the writer’s knowledge, only a couple of studies have been conducted in this research area. Polynomial modeling of power amplifiers includes basic polynomial models, Volterra series, memory polynomials (MP), generalized memory polynomials (GMP), Wiener model, Hammerstein model and parallel Hammerstein model. The parameters for polynomial models can be estimated with help of the least squares (LS) method.
Machine Learning (ML) is a component of artificial intelligence (AI), which enables systems to automatically learn from prior experiences without the programmer. ML algorithms are often grouped by similarity so that algorithms, which work in the same way, are collected together. Usually, they are also categorized as supervised and unsupervised. Supervised algorithms may apply what has been learned earlier to new data using labeled examples to estimate upcoming events. In contrary, unsupervised algorithms are used when the training data is neither classified nor labeled. ML algorithms include, for example, regression algorithms, instance-based algorithms, regularization algorithms, decision tree algorithms, Bayesian algorithms, artificial neural network algorithms and ensemble algorithms.
The performance of various polynomial and machine learning techniques were evaluated with NMSE and ACEPR values. NMSE was utilized in both techniques. In the experiments, the MATLAB interface was used. The utilized techniques were Linear Regression, a two-layer feed forward neural network, a couple of Support Vector Machine methods, Decision Trees (including Fine Tree, Medium Tree and Coarse Tree), memoryless polynomial model and memory polynomial model.
The memoryless polynomial had similar NMSE as ML techniques, but decision trees, especially Fine Tree method, seemed to perform better than other ML techniques. The memory polynomial had better performance than the memoryless one.
The DTs had the best performance, because the DTs have several advantages compared to other techniques. For example, DTs perform rather well with large sets of data, they may process both numerical and categorical data demanding only little data preparation. These aspects may be the reasons why Fine Tree had the smallest NMSE compared to other ML methods.
Power amplifier modeling is classically conducted using polynomials. It can also be done using machine learning. However, modeling them with machine learning is much more uncommon. According to the writer’s knowledge, only a couple of studies have been conducted in this research area. Polynomial modeling of power amplifiers includes basic polynomial models, Volterra series, memory polynomials (MP), generalized memory polynomials (GMP), Wiener model, Hammerstein model and parallel Hammerstein model. The parameters for polynomial models can be estimated with help of the least squares (LS) method.
Machine Learning (ML) is a component of artificial intelligence (AI), which enables systems to automatically learn from prior experiences without the programmer. ML algorithms are often grouped by similarity so that algorithms, which work in the same way, are collected together. Usually, they are also categorized as supervised and unsupervised. Supervised algorithms may apply what has been learned earlier to new data using labeled examples to estimate upcoming events. In contrary, unsupervised algorithms are used when the training data is neither classified nor labeled. ML algorithms include, for example, regression algorithms, instance-based algorithms, regularization algorithms, decision tree algorithms, Bayesian algorithms, artificial neural network algorithms and ensemble algorithms.
The performance of various polynomial and machine learning techniques were evaluated with NMSE and ACEPR values. NMSE was utilized in both techniques. In the experiments, the MATLAB interface was used. The utilized techniques were Linear Regression, a two-layer feed forward neural network, a couple of Support Vector Machine methods, Decision Trees (including Fine Tree, Medium Tree and Coarse Tree), memoryless polynomial model and memory polynomial model.
The memoryless polynomial had similar NMSE as ML techniques, but decision trees, especially Fine Tree method, seemed to perform better than other ML techniques. The memory polynomial had better performance than the memoryless one.
The DTs had the best performance, because the DTs have several advantages compared to other techniques. For example, DTs perform rather well with large sets of data, they may process both numerical and categorical data demanding only little data preparation. These aspects may be the reasons why Fine Tree had the smallest NMSE compared to other ML methods.