An Incremental Learning Structure Using Granular Computing and Model Fusion with Application to Materials Processing
This chapter introduces a Neural-Fuzzy (NF) modelling structure for offline incremental learning. Using a hybrid model-updating algorithm (supervised/unsupervised), this NF structure has the ability to adapt in an additive way to new input–output mappings
- PDF / 629,540 Bytes
- 15 Pages / 439.18 x 665.994 pts Page_size
- 66 Downloads / 171 Views
tion Data driven Computational Intelligence (CI) models are often employed to describe processes and solve engineering problems with great success. NeuralNetworks (NN) [7, 11], Fuzzy, and Neural-Fuzzy systems (NF) [3, 10] as well as Evolutionary and Genetic Algorithms (GA) [9,12] have all been used in the past to solve real-world modelling and control engineering problems. The scientific maturity of such methodologies and the demand for realistic representations of high complexity real-world engineering processes allowed new and advanced features of such systems to evolve. Such features include the ability of a model/structure to incrementally learn from new data. These structures are able to learn from an initial database (with appropriate training) but at the same time incrementally adapt to new data when these are available. Additional requirements include the system’s ability to interact with the environment in a continuous fashion (i.e. life-long learning mode) and having an open structure organisation (i.e. dynamically create new modules) [5]. Several G. Panoutsos and M. Mahfouf: An incremental Learning Structure Using Granular Computing and Model Fusion with Application to Materials Processing, Studies in Computational Intelligence (SCI) 109, 139–153 (2008) c Springer-Verlag Berlin Heidelberg 2008 www.springerlink.com
140
G. Panoutsos and M. Mahfouf
methods have been developed so far that demonstrate some of the aspects of incremental learning systems [2, 4–6]. This chapter presents a new approach that is based on Granular Computing Neural-Fuzzy networks (GrC-NF) [10]. By using such an approach, it is possible to achieve a high single network performance and at the same time maintain a high system transparency (in contrast to black-box low transparency modelling i.e. NN). An incremental learning architecture is subsequently developed using the GrC-NF models in a cascade model-fusion manner. The entropy of the cascade models is used as the main feature that assists the data fusion process. The proposed methodology is tested against a multidimensional MISO real industrial application; the prediction of mechanical properties of heattreated steel is investigated. Such application involves complex databases, containing data with non-linear dynamics and high interaction between the dimensions as well as sparse data of high uncertainty (measurement noise, operator errors, etc.).
2 GrC – NF Modelling Models elicited via the Granular Computing – Neural-Fuzzy (GrC-NF) approach, as described in [10], will form the main building blocks of the Incremental Learning (IL) system. This modelling process is realised in three steps: 1. Knowledge discovery 2. Rule-base creation 3. Model optimisation 2.1 Knowledge Discovery Granular Computing mimics the perception and the societal instinct of humans when grouping similar items together. Grouping items together is not just a matter of proximity but also of similarity measures such as similarity in class or function. Data granulation [1, 9] is achieved by a simple and transpar
Data Loading...
