No description available.

No description available.

Computer science literature abounds with descriptions of program verifiers, systems which analyse a software program and attempt to prove automatically that the program satisfies behavioural specifications. Techniques used include predicate abstraction, three-valued heaps graphs and classes of polyhedra. Yet while these systems have had some encouraging successes, each deals only with particular patterns of program behaviour: e.g. predicate abstraction can infer arithmetical relationships but does not capture the 'shape' of linked data structures; for three-valued shape graphs the reverse is true. Thus typical programs, in which different patterns of behaviour are mixed up together, still cannot be verified automatically. This thesis explores the question: 'By combining several program verifiers, and making them cooperate, can we produce a verification system that solves a broader range of verification problems than its components do?'. Specifically, our approach is to allow the verifiers to exchange information about program states, expressed as formulae of a single common logic, so that each can benefit from the others' findings. We design a mechanism which enables the verifiers to cooperate. Our setup comprises several verifiers, cleanly separated as analysis modules implementing a common interface, and a central 'broker' which overseas the verification process, propagating formulae between the analysis modules. We formalise this approach for programs of a core imperative heap-manipulating language with recursion, annotated with correctness assertions. We give an interprocedural verification algorithm for the broker and soundness conditions for analysis modules, and prove that these ensure the algorithm is sound (though perforce incomplete). We report on the implementation of our new method in an experimental system HECTOR, which includes the broker and analysis modules for a range of techniques, including predicate abstraction and three-valued shape analysis. By means of a verification case study, we demonstrate some of the advantages of our approach.

No description available.

Three stepsize strategies are considered for Sequential Quadratic Programming and Primal-Dual Interior Point methods. Each strategy has its advantages and disadvantages. Switch rules are implemented, that combine all three strategies so as to complement their merits and avoid their drawbacks. Both algorithms are proven to converge to a local minimum from arbitrary starting points, yielding gl.obal convergence. In addition, it is shown that unity stepsizes are achieved. Consequently, superlinear convergence is maintained. Also, the penalty parameter is updated finitely often, which thereby avoids numerical difficulties. Promising results on small and larger scale problems validate the global and superlinear convergence properties of the algorithms and also display that the value of the penalty parameter is kept low at all times. Finally, the algorithms are used as building blocks for solving Mixed Integer Nonlinear Programming probl~ms using the Generalized Benders Decomposition and the Generalized Outer Approximation.

No description available.

No description available.

No description available.

This thesis proposes a new methodology for evaluation of automatic music genre classification. It is argued that the common tacit understanding that genre is an attribute of a piece is unsound, and that genre is better understood as an attribute given to a piece by a group of people. As a direct consequence of this, different groups will assign different genre labels to the same piece.

No description available.