Date of Submission


Date of Award


Institute Name (Publisher)

Indian Statistical Institute

Document Type

Doctoral Thesis

Degree Name

Doctor of Philosophy

Subject Name

Computer Science


Electronics and Communication Sciences Unit (ECSU-Kolkata)


Pal, Nikhil Ranjan (ECSU-Kolkata; ISI)

Abstract (Summary of the Work)

There are problems of interest for which precise mathematical or physical understanding is yet to come. Let us illustrate this with an example. Suppose, for a mining operation one needs to blast a certain portion of the soil/rock using some specific ex- plosives. Before the blast is made, the miners want to know the intensity of vibration that would be produced at a certain distance from the site of blasting. The intensity of vibration may de pend on several factors. For example, it will depend on the characteristics and quantity of the explosive used, the rock characteristics of the region, the distance between the blasting site and the point where vibration is measured. But, we do not know precisely how these factors control the intensity of vibration. There are physics based models which model the grosscenario, but they are based on assumptions which are sometimes too simple to be useful. There are numerous such problems for which we do not yet have precisephysics basedmodels. For such problems, it is known that the out put s bear certain relationship with the possible inputs. But the un derlying process governing the relation between the inputs and outputs is believed to be so complex that it is difficult to formulate physical models to explain it to our satisfaction. For such problems it is possible to obtain past data in the form of input- out put observations We can assume that there is some unknown function which maps the relevant in put space to the output space. Based on the available in put-out put data our task is to construct a computation analyst em which can act as the function to per- form this transformation. Such systems built from data attempt to mimic the original system, and can be used for future predictions. This thesis deals with identification of such systems from input-output data. Systems which have been developed using in put-out put examples are sometimes called learningsy systems Such systems implicitly or explicitly learn some function from data These can be classified into two categories based on the type of out put they produce. Systems in the first category produce numerical value s as outputs (these are termed as аге function approximation type systems). For example, one may train a system to predict the price of a stock based on available data. Such a system will produce a numerical output, which in this case is the predicted price of a stock. Systems of the second kind do not produce numerical values but class labels or some decision (such a system can be called a classifier type system). For example, a classifier can be designed to decide whether a pixel denotes land or water from the gray level characteristic of a remotely sensed image.


ProQuest Collection ID:

Control Number


Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.


Included in

Mathematics Commons