Please use this identifier to cite or link to this item:
|Title:||Τεχνητή νοημοσύνη: ασαφείς γνωστικές απεικονίσεις και νευρώνες βεβαιότητας.|
|Alternative Titles:||Artificial intelligence: fuzzy cognitive maps and certainty neurons.|
|Department:||Πανεπιστήμιο Μακεδονίας Οικονομικών και Κοινωνικών Επιστημών. Τμήμα Εφαρμοσμένης Πληροφορικής (ΕΠ)|
|Abstract:||In this dissertation, the Certainty Neuron (CN) is proposed as new type of artificial neuron that uses a two variable transfer function and is augmented with local memory and decay mechanisms. The first variable of CN transfer function is the one used by classical artificial neurons that is the weighted sum of influences that the neuron receives from all other neurons. The second variable is the current activation lever of the neuron. This means that CN have memory of its previous output and its new activation level does not depend only on the weighted sum of influences that it receives from other neuron but also on its previous activation level. The function that CN uses to aggregate these two values is function fM that was used for the handling of certainty factors in the MYCIN expert system. Our analysis of function fM show that it is a uninorm operator and so it is suitable for the aggregation of values coming from the quantification of linguistic variables. Furthermore it was found that this function can be considered as an extension to the three dimensional space, of the common step, linear and sigmoid function that are used as threshold functions in artificial neurons. The CN transfer function also employs a decay mechanism that subtracts from the new activation a part of the previous activation level as an act of the natural neuron’s tendency to become completely not activated. This also imposes the notion of time to the CN structure. CN are proposed as neurons that possess increased inference capabilities when used in the Fuzzy Cognitive Map (FCM) structure. FCM is an extension of Cognitive Map (CM) used mainly for Decision Making and Prediction. It is an Artificial Neural Network creating models by using neurons to represent the concepts of the model and the weights of the network’s connections to represent the strength of the causal relationships between the concepts of the model. After a solid introduction to CM and FCM, we study the elementary two Certainty Neuron FCM (CNFCM) and we see that in the case of positive feed back it is more stable that the corresponding two neuron FCMs. In the case of negative feedback, the two neuron CNFCM show that it has enhanced representing capabilities by exhibiting limit cycle behaviour of period that depends on the parameters of the CNFCM structure while corresponding FCM structure exhibits limit cycle of period 4, regardless the FCM parameters. The inference capabilities of CNFCM structure or more neurons, are presented and compared with that of classical FCM structures. CNFCM structure show increased inference capabilities by being capable to estimate not only the increase or decrease of each model’s concept but also the degree of this increase or decrease. The CNFCM structure was also found suitable to be used for strategic planning. The Dynamical Behaviour of CNFCM was studied by a series of simulations. Having as control parameters the neuron’s decay factor and the symmetry of the FCM weight matrix, we calculate the entropy of the simulated systems in order to draw conclusions about their dynamical behaviour. Specific areas in the parameter space were found where CNFCM exhibits limit cycle behaviour, stable fixed point behaviour or collapse. Our study show that the increase of the decay factor can cause a CNFCM to change from limit cycle behaviour to a stable fixed point behaviour by the decrease of the limit cycle magnitude. Furthermore it was found that as the CNFCM weight matrix becomes more antisymmetric, the fixed point behaviour changes to limit cycle behaviour by the creation of limit cycles with very big period. The dynamical behaviour of CNFCM was compared with that of other FCM structures and found to be completely different from that of FCM structures using neurons not having memory, like the sigmoid FCMs.|
|Keywords:||Ασαφείς γνωστικές απεικονίσεις|
Μη γραμμική δυναμική
Δίκτυα, τεχνητά νευρωνικά
|Information:||Η βιβλιοθήκη διαθέτει αντίτυπο της διατριβής σε έντυπη μορφή.|
Διατριβή (Διδακτορική)--Πανεπιστήμιο Μακεδονίας, Θεσσαλονίκη, 1997.
Περιλαμβάνει βιβλιογραφικές αναφορές (σ. 222-231).
|Rights:||Το ψηφιακό τεκμήριο της διατριβής αποτελεί παραχώρηση του Εθνικού Αρχείου Διδακτορικών Διατριβών που τηρεί το Εθνικό Κέντρο Τεκμηρίωσης σύμφωνα με το αρ. 22 του Ν. 2121/1993|
|Appears in Collections:||Τμήμα Εφαρμοσμένης Πληροφορικής (Δ)|
Files in This Item:
|TsadirasAthanasios_Phd1997.pdf||17.17 MB||Adobe PDF||View/Open|
Items in Psepheda are protected by copyright, with all rights reserved, unless otherwise indicated.