Main constituents: - Fuzzy systems - Neural networks - Evolutionary computing - Probabilistic reasoning VAN-00 2 Constituents of SC Fuzzy systems => imprecision Neural networks => learning Probabilistic reasoning => uncertainty Evolutionary computing => optimization Over 24 000 publications today VAN-00 3 SC: a user-friendly interface Soft computing approach Linguistic world Soft data Interpretations Understanding Explanations Qualitative methods Bivalent or multivalent reasoning Mathematical world Hard data Quantitative methods Bivalent reasoning Phenomenon under study VAN-00 4 Advantages of SC Models base on human reasoning. Models can be - linguistic - simple (no number crunching), - comprehensible (no black boxes), - fast when computing, - good in practice. VAN-00 5 SC today (Zadeh) Computing with words (CW) Theory of information granulation (TFIG) Computational theory of perceptions (CTP) VAN-00 6 Possible SC data & operations Numeric data: 5, about 5, 5 to 6, about 5 to 6 Linguistic data: cheap, very big, not high, medium or bad Functions & relations: f(x), about f(x), fairly similar, much greater VAN-00 7 Neural networks (NN, 1940's) Neural networks offer a powerful method to explore, classify, and identify patterns in data. Website of Matlab Neuron: y=w i x i
Inputs Neurons (1 layer) Outputs VAN-00 8 Machine learning (supervised) Pattern recognition based on training data. Classification supervised by instructor. Neural (crisp or fuzzy), neuro-fuzzy and fuzzy models. Peach Plum ? Instructor VAN-00 9 Machine learning (unsupervised) Pattern recognition based on training data. Classification based on structure of data (clustering). Neural (crisp or fuzzy), neuro-fuzzy and fuzzy models. Peach Plum Nectarine Labeling VAN-00 10 Machine learning (unsupervised) Self-organized maps (Kohonen). Fuzzy c-means (Bezdek). Subclust (Yager, Chiu). Peach Plum Nectarine Labeling Websom Self-Organizing Maps for Internet Exploration VAN-00 11 Fuzzy systems (Zadeh, 1960's) Deal with imprecise entities in automated environments (computer environments) Base on fuzzy set theory and fuzzy logic. Most applications in control and decision making Omrons fuzzy processor Omron Electronics Matlab's Fuzzy Logic Toolbox VAN-00 12 SC applications: control Heavy industry (Matsushita, Siemens, Stora-Enso) Home appliances (Canon, Sony, Goldstar, Siemens) Automobiles (Nissan, Mitsubishi, Daimler- Chrysler, BMW, Volkswagen) Spacecrafts (NASA) VAN-00 13 SC applications: business hospital stay prediction, TV commercial slot evaluation, address matching, fuzzy cluster analysis, sales prognosis for mail order house, multi-criteria optimization etc. (source: FuzzyTech) supplier evaluation for sample testing, customer targeting, sequencing, scheduling, optimizing R&D projects, knowledge-based prognosis, fuzzy data analysis VAN-00 14 SC applications: finance Fuzzy scoring for mortgage applicants, creditworthiness assessment, fuzzy-enhanced score card for lease risk assessment, risk profile analysis, insurance fraud detection, cash supply optimization, foreign exchange trading, insider trading surveillance, investor classification etc. Source: FuzzyTech
VAN-00 15 SC applications: robotics
Fukudas lab Joseph F. Engelberger We are proud to announce that the HelpMate Robotic Courier has been acquired by Pyxis Corporation. Entertainment robot AIBO VAN-00 16 SC applications: others
Statistics Social sciences Behavioural sciences Biology Medicine VAN-00 17 (Neuro)-fuzzy system construction Training data Experts Fuzzy rules (SOM, c-means etc.) Control data System evaluation (errors) Tuning (NN) New system VAN-00 18 0 0,2 0,4 0,6 0,8 1 1,2 0 2 4 6 8 10 12 X Y Model construction (mathematical) Mathematical models are functions. Deep knowledge on mathematics. If non-linear (eg. NN), laborious calculations and computing. Linear models can be too simplified. How can we find appropriate functions? Y=1-1./(1 + EXP(-2*(X-5))) VAN-00 19 Model construction (trad. rules ) If 0<x<1, then y=1 If 1<x<2, then y=0.99 : If 8<x<10, then y=0 If 0<x<1, then y=f(x) If 1<x<2, then y=g(x) : If 8<x<10, then y=h(x) 0 0,2 0,4 0,6 0,8 1 1,2 0 2 4 6 8 10 12 X Y - Rule for each input. => Large rule bases. - Only one rule is fired for each input. - Coarse models. VAN-00 20 0 0,2 0,4 0,6 0,8 1 1,2 0 2 4 6 8 10 12 X Y Model construction (SC/fuzzy) If x0, then y1 If x5, then y0.5 If x10, then y0 - Approximate values - Rules only describe typical cases (no rule for each input). => Small rule bases. - A group of rules are partially fired simultaneously. VAN-00 21 SC and future SC and conventional methods should be used in combination.