Chapter-2 motivation -> bf or subsubsection Beyes error is however difficult => Beyes error is, however, difficult or However Beyes error is difficul alphabet with different font is a good idea ============================================================= [Structure of the thesis] [At the beginning of chapter 2] the term complexity estimation is wider than the one defined here such as Kolomogolov. Where is the definition of complexity estimation described explicitly? According to the conclusion section of chapter 2, 23 different ways of how we estimate the complexity are presented. I guess they are in 2.4.3 "Mesures related to space partitioning" and 2.4.4 "Other Mesures". But IMHO there is no explicit definition of complexity estimation. I think the overall structure of chapter 2 SHOULD be "Complexity estimation is defined as ... where .............." For example, author says at the top of p.68 that "In order to measure complexity, CDM uses ..." and CDM is difeined. But how complexity is mesured or how complexity relates to the CDM is never mentioned in 2.4.3.1. The same is true for 2.4.3.2. 1) probability of class i in cell l 2) degree of separability 3) purity are given but no description of how they are used to know the complexity. Frastrating, isn't it? Thus from 2.4.3.1 to 2.4.3.4 and 2.4.4.1 to 2.4.4.8 various mesures are described but no description of how each of them related to complexity. From what you wrote in p.71 that "The main drawback is that it acts more like a detector of linearly separable classes than complexity measure" all of those mesures might be used to estimate the complexity as it is. And finally I find the description in p.74 that "Average proportional size in terms of number of members in each hypersphere divided by the total number of data points is used as a measure of complexity." And in 2.4.5 you propose using several mesures to estimate complexity. Hence I'm sure those mesures are all proposed to express the mesure of complexity. But in onrder to combine those mesures, IMHO, it is necessary normalize them. you might include the normalization factor into the weights, but weights is essentially to express importance of each measure. [chapter 3] The structure of each chapter should be consistent. from this aspect "General description of T-DTS" and "Introduction - Modular algorithms" should be sections or subsections. For example 3.1 General description of T-DTS 3.2 Introduction - Modular algorithms koko ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [Just a matter of preference but... ] reference [Style] do not use bf in the text [need to be in the list of refer] p.83 Mixture of Experts (who started to use this term?) ref pier oho -> ohio [MADA03A] conference name? two [HO 00] exist p.50 Basu is in [HO 00] and [HO 02] not [HO 01] [strange terminology but liegal] hypercuboids. modelization ======= chapter 3 -> chaptemotivation research p.49 motivation -> no need There is a great impact of Bayes error on the most of classification complexity techniques. => Bayes error relates to the most of classification complexity techniques. p.51 S as function of T => S as a function of T estimated by model f => estimated by the model p.53 all input values for whose output is ci => all input values whose output is ci. p.55 however -> ? only title no explanation e.g. naive Bayes (see e.g.) like p.58 p.59 Regarding that fact => Therefore this section present in detail classification complexity estimation methods => this section explains classification complexity estimation methods in detail categorized as stated in previous section. => no need. Delete! Section 2.4.5 talks => describes creating ensembles of estimators => exploiting ensemble of estimators which correctly used can inhibit advantages over solo methods. => which can exhibit advantages over solo method if exploited correctly. p.65 relay on -> rely on? p.65 where ... should be no indended p.66 writes -> wrote PhD -> Ph.D. in his partial phd work -> ? with a much more simple computational complexity. -> oxymoran where => noindent in his partial => second time --- p. 67 --- 2.4.3 Measures related to space partitioning => 2.4.3 Measures Related to Space Partitioning following => the following That allow to obtain => That allows us to obtain --- p.68 --- In order to measure complexity, CDM uses only Not Linearly Separable Terminal Boxes, as, according to author, only these contribute to Bayes error. That is however not true, => In order to measure complexity, CDM uses only Not Linearly Separable Terminal Boxes. The author explained that this is because only they contribute to Bayes error. However this is not true. cannot by itself diminish the Bayes error of the whole dataset, => cannot diminish the Bayes error of the whole dataset by itself, it can help classifiers in approaching the Bayes error optimum. => it can help classifiers approach the optimum Bayes error. The formula for CDM is => The CDM is difined as where k(i) => Do not indent For task that lead => For tasks that lead koko like Neighbourhood Separability [SING 02A] measure is developed by Singh. Similarly over indended like p.69 The Fisher discriminant ratio => izen nakatta? The advantage is very low computational complexity. -> oxymoran --- p.69 --- 1<=k<=il -> not a good way --- p.72 --- in which classed don't overlap => in which classes don't overlap --- p.74 --- they frequently base uniquely on experimental data => they are frequently based on experimental data as they obtain that in experimental way during the processing. => as they obtain the distribution in experimental way during the processing. designed only to two-class problems => designed only for two-class problems what usually result in combinatorial increase in computational complexity. => which usually result in combinatorial increase in computing comlexity ~~~~~~ ~~~~~~~~~ --- p.75 --- autoadaptation of processing system to the data => autoadaptation of data processing system --- p.81 --- Motivation: This chapter has to present in detail the new Treelike-Divide To Simplify approach, define its structure, and describe the types of modules that are used in the structure. It will present also in detail procedures and algorithms that are used for the creation, execution and modification of modules. It will discuss also advantages and disadvantages of T-DTS approach and compare it with other approaches. => This chapter presents what is Treelike-Divide To Simplify (T-DTS) approach, defines its structure, and describes the types of modules that constructs the structure. Procedures and algorithms to create, execute and modify modules are also presented. Then advantages and disadvantages of T-DTS approach are discussed by comparing it with other approaches. In a very large number of cases, dealing with real world dilemmas and applications (system identification, industrial processes and manufacturing regulation and optimization, decision, pattern recognition, systems and plants safety, etc.), information is available as data stored in files (databases etc.). => When we deal with real world problems information is given as data files. Especially in industrial areas, efficient processing of data is the chief condition to solve problems. => Especially in industrial areas, efficient processing of data is requiered to solve problems. appropriated => appropriate (2 places in this page) allied parameters make complexity => ? --- p.82 --- "divide et impera" is the same as so called "divide and conquer"? T-DTS (Treelike Divide To Simplify) paradigm => T-DTS paradigm (just one time explanation of an abbrebiation is enough) allows parallel processing and simplification of the problem => allows us to simplify a problem and process them in parallel. The purpose is based on the use of a small set of specialized mapping Neural Networks, called Neural Network Models (NNM), supervised by a Decomposition Unit (DU). => We use a set of Neural Networks supervised by a Decomposition Unit (DU). Decomposition Unit could be a prototype based neural network, Markovian decision process, etc. => DU can be constructed using a neural network, Markovian decision process, etc. The modules responsible for processing in the structure are Artificial Neural Networks (models). The T-DTS paradigm allows us to build a tree structure. At the nodefs level, the input space is decomposed into a set of subspaces of smaller sizes. At the leaf's level the aim is to learn the relations between inputs and outputs relatives to one of sub-spaces, obtained from splitting. => The T-DTS paradigm builds a tree structure in which input space is decomposed into a set of subspaces of smaller sizes at the node's level, while at the leaf's level, the relations between inputs and outputs in the splitted sub-spaces are learned. --- Section 3.7 discuses properties => Section 3.7 discusses properties The approaches most close to presented in this work => The approaches most close to the ones presented in this work ~~~~~~~~ --- p.83 --- which splits recursively data => which recursively splits data Mixture of Experts -> add reference An issue could be model complexity reduction by splitting of a complex problem into a set of simpler problems: multi-modelling where a set of simple models is used to sculpt a complex behaviour [GOON 96]. Another promising approach to reduce complexity takes advantage from hybridization [KROG 95]. Several ANN based approaches were suggested allowing complexity and computing time reduction. Among proposed approaches, one can note the Intelligent Hybrids Systems [KROG 95], Neural Network Ensemble concept [HANN 93], Models or experts mixture ([BRUS 95], [SUNG 95]), Dynamic Cell Structure architecture [LANG 98] and active learning approaches [FAHL 90]. => One of the important issues is to reduce complexity by splitting a complex problem into a set of simpler problems: [GOON 96] proposed the multi-modelling method where a set of simple models is used to sculpt a complex behaviour; [KROG 95] reduced complexity by using hybridization and called it the Intelligent Hybrids Systems; Also ANN based approaches such as Neural Network Ensemble concept [HANN 93] were proposed to recuce complexity and computing time; Models of experts mixture ([BRUS 95], [SUNG 95]), Dynamic Cell Structure architecture [LANG 98] and Active Learning Approaches [FAHL 90]. were the other examples. There is a great variety of intelligent software agents and structures. T-DTS approach processing leaves (NN Models) can be classified as computational, software and task-specific agents. The agents are created, modified and directed by Decomposition Units (DU), creating a tree structure. => There are a great variety of intelligent software agents. T-DTS approach we propose can be classified as computational, software and task-specific agents. The agents are created, modified and directed by Decomposition Units (DU), creating a tree structure. One can see a strong resemblance to Distributed Artificial Intelligence (DAI) [WOOL 02] systems. --- p.86 --- General description of T-DTS etc -> numbered subsection p.87 Preprocessing phase -> should be explained as learning and operation phase what does it mean by (by dimension) p.93 however => , however, figure Figure 3.5 the location of the caption of the Fig 4.5, Fig 4.7, Fig.4.9 tree.The => tree. The The location of the caption for the Table 4.1 koko ===== from p.83