EVOlutionary Learning by Intelligent Variation and choice of suitable Operator sets

EVOLIVO - EVOlutionary Learning by Intelligent Variation and choice of suitable Operator sets – is the name of our project describing our research in the area of Evolutionary Algorithms and Programs including problems of analysis, improvement and on-line adaptation of existing standard operator sets (basically selection and recombination) and the development of new problem-specific operator sets and their application to problems of technical interest, in particular to automatic system synthesis. Recent research was focused on multi-objective optimization with uncertain objectives. Here, it is assumed that the objectives of a solution are not precisely known, but are given by intervals. These intervals determine the minimal and maximal values of an objective. An essential notion in multi-objective optimization is the term dominance describing the superiority of a solution over another solution. In case of uncertain objectives the dominance is no longer defined. This obviously restricts the use of multi-objective optimization methods in the presence of fuzzy objectives. In the EVOLIVO project we succeeded to generalize these optimization strategies by defining the so-called probability of dominance regarding uncertain objectives. These results are expected to have impact on other research areas outside the domain of embedded systems. Last year we successfully applied these multi-objective optimization strategies to the task of automatic design space exploration. Therein we proposed methods for hierarchical design space exploration using a class of so-called Pareto-Front-Arithmetics for the accelerated design space exploration. The idea of Pareto-Front-Arithmetics is that generally, optimization problems are of hierarchical nature and can be decomposed into sub-problems. By combining the partial results of the optimization, we have to consider the problem that, in general, a global optimum is not composed of optima of its sub-problems. This only holds for monotonous objective functions. The contribution of this project is that we have shown the viability of the combination of partial results leading to uncertain objectives for the top-level optimization problem. By using the dominance probability, we separated the optimization problem and we were able to construct good approximations of the overall solutions in a short time. Future work will focus on the integration of population-based optimization methods into dynamic systems. We will analyze the usability of these methods for online optimization. Target architectures are networked embedded systems. Beside Evolutionary Algorithms we will explore novel population-based optimization strategies like Ant Colony Optimization and Particle Swarm Optimization.