- author = "Sunith Bandaru",
- title = "Automated Innovization: Knowledge discovery through multi-objective optimization",
- school = "Indian Institute of Technology Kanpur",
- year = "2013",
- address = "India",
- keywords = "genetic algorithms, genetic programming, Innovization, MEMS",
-
URL = "
https://drive.google.com/file/d/0B8WHZC_8VuhxZ3FWenBfa19MSDQ/view",
-
URL = "https://www.iitk.ac.in/kangal/deb_phd.shtml",
- size = "227 pages",
-
abstract = "In recent years, there has been a growing interest in
the field of post-optimality analysis. In a single
objective scenario, this usually concerns optimality,
sensitivity and robustness studies on the obtained
solution. Multi-objective optimization on the other
hand, poses an additional challenge in that there are a
multitude of possible solutions (when the objectives
are conflicting) which are all said to be
Pareto-optimal. These solutions may collectively hold
crucial design information. Properties that are common
to all or most Pareto-optimal solutions can be
considered as characteristic features that make good
designs. Knowledge of such features, in addition to
providing better insights into the problem at hand,
enables the designer to hand craft solutions for other
optimization tasks which are structurally similar to
it; thus eliminating the need to actually optimize.
Innovization is the process of extracting this
information in the form of design principles, which are
basically mathematical expressions relating various
problem entities like variables, objectives and
constraint functions. Manual innovization involves the
visual identification of correlations between problem
entities through two and three-dimensional plots.
Thereafter, appropriate functions are used for
regression and the design principles are obtained.
Though the procedure has been applied to many
engineering design problems, the human element involved
in it limits its use in most practical
applications.
The present thesis firstly proposes automated innovization, an unsupervised machine learning technique that can identify correlations in any multi-dimensional space formed by variables, objectives, etc. specified by the user and subsequently performs a selective regression on the correlated part of the Pareto-optimal dataset to obtain a design principle. The correlations are automatically identified by a customized grid-based clustering algorithm and the design principle is evolved using a genetic algorithm. Next, the procedure is extended so that design principles hidden in all possible Euclidean spaces formed by the variables and objectives (and any other user-defined functions) can be obtained simultaneously, without any human interaction, in a single run of the algorithm. This is accomplished by introducing a niching strategy to evolve different species of design principles in the same population of a genetic algorithm. Automation in innovization is achieved at the cost of restricting the mathematical structure of the design principles to a certain form, the significance of which becomes clear by observing physical laws in nature. Later in this thesis, a tree-based genetic programming framework is integrated into automated innovization to obtain design principles of any generic mathematical structure. Dimensionality information is introduced in the search process to produce design principles that are meaningful to the designer.
Next, the proposed automated innovization technique is used to obtain design principles for four real-world multi-objective design optimization problems from varied fields. They are: noise barrier design optimization, polymer extrusion process optimization, friction stir welding process optimization and MEMS (MicroElectroMechanical Systems) resonator design optimization. In each case the obtained design principles are presented to experts of the respective fields for interpretation to gain insights.
Secondly, this thesis introduces two new innovization concepts, namely higher-level innovization and lower-level innovization. Multi-objective optimization problem formulations involve many settings that are not changed during the solution process. However, once the trade-off front is generated, the designer may wish to change them and rerun the optimization, thus obtaining more fronts. This happens in many real-world situations where the designer is initially unsure about problem elements such as constraints, variable bounds, parameters and even objective functions. Higher-level innovization answers questions like: Are the features of the original problem still valid for other generated fronts? If not, how do they change with the modified setting?. The name reflects the fact that higher-level design knowledge is gained in the process. Sometimes lower-level design knowledge may also be desired. Consider the situation when after obtaining a set of trade-off solutions for a multi-objective design problem, a posteriori decision-making approach is used to identify a region of preference on the trade-off front. Now the designer may be interested in knowing features that are common to solutions only in this partial set and are not seen in rest of the trade-off solutions, so that the designer is specifically aware of properties associated with the chosen solutions. In this thesis, the automated innovization technique is extended to perform both higher and lower-level innovization.
Thirdly, this thesis studies the temporal evolution of design principles obtained using automated innovization during the course of optimization. Results on a few engineering design problems reveal that certain important design features start to evolve early on, whereas some detailed design features appear later during optimization. Interestingly, there exists a simile between evolution of design principles in engineering and human evolution. Such information about the hierarchy of key design principles should enable designers to have a deeper understanding of their problems.",
- notes = "SunithBandaru_PhDThesis.pdf",
