Parallel algorithms in design space exploration

OData support
Dr. Horváth Ákos
Department of Measurement and Information Systems

The goal of Design Space Exploration (DSE) is to semi-automatically synthesize various design candidates satisfying numerical and structural constraints. DSE is frequently used in Model Driven Systems Design (MDSD) to partially automate design processes or to dynamically reconfigure autonomous systems. Rule-Rased DSE aims to achieve this by starting from an initial model which evolves by transformation rules until the desired goals are reached (which are captured by model queries). As a result, rule-based DSE finds a sequence of rules which transforms the initial model to reach a target state satisfying the goals.

Many DSE challenges can be seen as Multi-Objective Optimization (MOO) problems i.e. it should find valid solutions while maximizing or minimizing several numerical values derived from the model. Genetic Algorithms (GA) and other meta-heuristics are widespreadly used for MOO. They maintain a predefined number of solutions or individuals and iteratively create new ones by mutation and crossover operations while it drops the low quality candidates.

In this report I present an approach (developed in collaboration with researchers from BME-MIT and Université de Montréal) to exploit multi-objective optimization techniques and use genetic algorithms for rule-based design space exploration keeping both domain independence and high level abstraction. Individuals are represented as rule trajectories, mutation operations modify the trajectory, crossover operations exchange rules between trajectories, while objectives are defined by model queries or are derived from rule executions. The selection operator is based on the Non-dominated Sorting Genetic Algorithm (NSGA-II).

I developed a prototype implementation built upon the VIATRA-DSE framework which uses Eclipse-based tools like EMF for model representation and EMF-IncQuery for model queries. The newly created framework supports extensive configurations like different stop conditions, custom genetic operators and multithreaded execution. The practical feasibility of the approach is demonstrated by experimental evaluation carried out on two case studies from different application domains.


Please sign in to download the files of this thesis.