Friday, September 18, 2009
Lingjia Tang
Chair: Mary Lou Soffa
Advisor: Paul Reynolds; Worthy Martin; Westley Weimer; Ron Williams
Rodman Room, 03:00:00
A Ph.D. Proposal
Combining Automated Reasoning and Search for Adaptation of Coercible Software
ABSTRACT
Adapting software to meet new requirements is highly desirable but its automation remains largely unachieved due to the intensive human effort required. The work proposed here will focus on the adaptation of a class of software we call coercible software, which is characterized by design flexibility in the model on which the software is based. Model design flexibility generally reflects subject matter expert uncertainty about model characteristics. Such uncertainty often provides opportunities for exploration of both structural and parametric alternatives in the model and its corresponding software. Adapting the software in accordance with expert-identified model alternatives can reveal new software behavior including that which corresponds to desired outcomes. A representative example of coercible software is simulations that are used for gaining insight into the phenomena they are modeling. Simulations used for gaining insight are the primary focus of the proposed research, although research results should apply to any software deemed coercible.
In those software adaptation cases where the distance between current software behavior and target behavior is measurable, some automation can be introduced by casting the problem as a search process. The search space can be constructed based on a subject matter expert~Rs knowledge about model assumption uncertainties and alternatives. Unfortunately, the search space typically grows exponentially. To address the issue of search space growth, and the broader issue of adaptation efficiency, the approach proposed here will combine automated reasoning and automated search methods, where the reasoning system is customized for the adaptation task. The research result will be an end-to-end framework that facilitates iterative semi-automated coercible software adaptation. The proposed framework includes three major components: (1) A knowledge representation system that uses Semi-Quantitative Differential Equations (SQDEs) to formally represent the subject matter expert~Rs knowledge about interrelations and parametric/structural uncertainties within the underlying model of the target software. (2) An automated reasoning system that navigates through search space and effectively prunes search space and guides the search process, using SQUID (Kay et al. 2000) as a component of predicting the software behavior based on SQDE and reporting the discrepancy between the predicted behavior and the user~Rs new software behavior requirements. (3) Dynamic knowledge incorporation that extracts knowledge from partial results of search as search proceeds and incorporates such newly gained knowledge into the reasoning system to guide further search agilely. The novelty of the proposed research lies in 1) its use of a semi-quantitative knowledge representation system to predict whether the software behavior will match desired behavior and 2) its combination of search space pruning with numerical optimization and dynamic knowledge incorporation to address the adaptation of coercible software. With success, my results will reduce the manual effort needed for adaptation and improve the efficiency of the adaptation process.