- Title
- A machine learning-based fuzzy framework for prediction problems
- Creator
- Fan, Zongwen
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2021
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- The rapid development of artificial intelligence, in terms of hardware implementation and software algorithms, has led to the widespread use of numerous prediction models, such as those related to finance, engineering and healthcare, in daily life. These models enable saving on labour and financial resources by making predictions about future outcomes based on historical data. Using extant prediction models, decision-makers can find and exploit patterns in data to detect risks and save time and money. However, the accurate prediction of outcomes in practical applications is very difficult because of the complexity, uncertainty and nonlinearity behind these real-world applications. To build a prediction model, traditional mathematical models and non-traditional methods (e.g. based on machine learning and computational intelligence models) can be used. In theory, machine learning models can provide alternate solutions with high approximation ability based on collected data by building prediction models to process real-world data and accurately analyse data and predict results. In reality, though, many prediction applications are difficult to formulate using precise mathematical models, owing to the difficulty of quantifying descriptive features into numerical values. Fuzzy systems, which apply a membership function to determine the degree of memberships, can represent degrees of truthfulness and falsehood instead of simple true and false values. They are able to transform descriptive data into a set of fuzzy rules, by using experts' experience and knowledge, thereby simulating human thinking with high interpretability. With higher model interpretability, decision-makers are more confident in making decisions since they understand the explanations about predictions. In this research, to use the high approximation ability of machine learning models and the high interpretability of fuzzy systems, these two have been combined to balance the models' approximation ability and interpretability by proposing a fuzzy framework based on machine learning, for considering prediction problems. In terms of fuzzy systems, this research focused on improving the construction of these systems as well as the fuzzy rule base. To improve the construction of the fuzzy model, it explored a multi-layer structure of fuzzy systems based on certain clustering methods, given that these methods can divide the data into groups based on the similarity of samples. By doing so, each cluster can be further divided into small groups using the same or different clustering methods. For improving the fuzzy rule base, this research investigated the use of machine learning algorithms to enhance the approximation ability of fuzzy rules. To validate the effectiveness of the multi-layer fuzzy framework based on fuzzy-rule clustering, first, a two-layer fuzzy system was proposed. Through experimental results from benchmark functions and datasets, the proposed two-layer fuzzy framework was found to exhibit high performance for prediction problems. Then, the two-layer fuzzy framework was extended to a multi-layer one. In addition, this research integrated machine learning models into fuzzy rules to further improve their approximation ability. Experimental results from a range of engineering benchmark problems confirmed that the proposed multi-layer fuzzy model can outperform other well-established fuzzy models in terms of accuracy and robustness without sacrificing efficiency. However, for bigger real-world problems with higher dimensions, such as the reverse prediction of concrete components in engineering and body fat prediction in healthcare, it was found that the proposed multi-layer framework does not work very well. Therefore, this research proposed the use of novel fuzzy-weighted machine learning approaches for the reverse prediction of concrete components and for body fat prediction. Specifically, a fuzzy operation was designed to alleviate the impact of noise data for machine learning models in different ways (in the model's error constraints or samples), considering that real-world data typically contain noise or outliers and current machine learning-based models are, to some extent, sensitive to noise in data. Experimental results confirmed that the proposed model performs very well for the reverse prediction of concrete components under both multi-input, one-output and multi-input, multi-output scenarios. This research also applied the proposed fuzzy-weighted approach to address a healthcare problem—body fat prediction. Experimental results showed that the proposed fuzzy-weighted model outperforms state-of-art models for body fat prediction.
- Subject
- artificial intelligence; machine learning; fuzzy framework; prediction problems; thesis by publication
- Identifier
- http://hdl.handle.net/1959.13/1502607
- Identifier
- uon:55247
- Rights
- Copyright 2021 Zongwen Fan
- Language
- eng
- Full Text
- Hits: 361
- Visitors: 388
- Downloads: 39
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | ATTACHMENT01 | Thesis | 20 MB | Adobe Acrobat PDF | View Details Download | ||
| View Details Download | ATTACHMENT02 | Abstract | 459 KB | Adobe Acrobat PDF | View Details Download |